<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: ML notes</title>
    <description>The latest articles on DEV Community by ML notes (@mlnotes_bio).</description>
    <link>https://dev.to/mlnotes_bio</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4257530%2F04f09708-2c91-455d-b4ad-dca6cc120b4a.png</url>
      <title>DEV Community: ML notes</title>
      <link>https://dev.to/mlnotes_bio</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/mlnotes_bio"/>
    <language>en</language>
    <item>
      <title>Segmenting budding yeast mother-bud pairs with SAM</title>
      <dc:creator>ML notes</dc:creator>
      <pubDate>Sat, 21 Nov 2026 21:19:59 +0000</pubDate>
      <link>https://dev.to/mlnotes_bio/segmenting-budding-yeast-mother-bud-pairs-with-sam-5h10</link>
      <guid>https://dev.to/mlnotes_bio/segmenting-budding-yeast-mother-bud-pairs-with-sam-5h10</guid>
      <description>&lt;h1&gt;
  
  
  Segmenting budding yeast mother-bud pairs with SAM
&lt;/h1&gt;

&lt;p&gt;Cell segmentation — defining cell outlines in microscopy images — is essential for quantitative image analysis. Budding yeast is a hard case: asymmetric cell division and mother-bud morphology cause dividing cells to be misidentified as two separate cells, which breaks downstream analysis. It is tempting to treat a general cell segmenter, or a vanilla Segment Anything Model (SAM), as good enough and to move on. This post stays inside one published study and asks a narrower question: when the error that matters is a split mother-bud pair, what does a yeast-specific SAM adaptation actually report?&lt;/p&gt;

&lt;p&gt;The paper is Zhao et al., &lt;em&gt;YeastSAM: A deep learning model for accurate segmentation of yeast cells&lt;/em&gt;, &lt;em&gt;Molecular Biology of the Cell&lt;/em&gt; 37(6):mr5, 2026 (&lt;a href="https://doi.org/10.1091/mbc.E25-09-0454" rel="noopener noreferrer"&gt;DOI 10.1091/mbc.E25-09-0454&lt;/a&gt;). The code is at &lt;a href="https://github.com/YonghaoZhao722/YeastSAM" rel="noopener noreferrer"&gt;https://github.com/YonghaoZhao722/YeastSAM&lt;/a&gt;, and the documentation is at &lt;a href="https://yeastsamdoc.readthedocs.io/" rel="noopener noreferrer"&gt;https://yeastsamdoc.readthedocs.io/&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  The setting, not a slogan
&lt;/h2&gt;

&lt;p&gt;YeastSAM is a deep learning segmentation framework for budding yeast, built by adapting SAM. The point of the adaptation is mother-bud morphology: dividing cells should be kept as correct instances rather than split into two. Well-separated, non-dividing cells may not show the same error, so the practical gain is largest on dividing populations.&lt;/p&gt;

&lt;h2&gt;
  
  
  What YeastSAM is
&lt;/h2&gt;

&lt;p&gt;The Crossref / journal abstract is direct: the authors addressed the mother-bud problem by adapting SAM to construct YeastSAM, a framework optimized for budding yeast. The open project is also a toolbox around that model. The GitHub repository and Read the Docs manual describe a GUI launcher and tools for generating masks, image registration, outline conversion, and mother-bud pair separation, written as a step-by-step guide for yeast segmentation and FISH-Quant integration. The documented tools support image series.&lt;/p&gt;

&lt;p&gt;In the documented workflow, masks are generated from DIC images, typically in napari with custom YeastSAM weights (&lt;code&gt;vit_b_yeast.pt&lt;/code&gt;, distributed on GitHub Releases). Those weights can also be used with BAND, the command-line interface, and QuPath. The docs recommend loading them rather than falling back to a default µSAM model. The repository acknowledges &lt;em&gt;Segment Anything for Microscopy&lt;/em&gt; (µSAM; Archit et al., &lt;em&gt;Nature Methods&lt;/em&gt;, 2025) as the inspiration; the published abstract describes YeastSAM as a SAM adaptation, and the open implementation is used with that microscopy SAM stack.&lt;/p&gt;

&lt;p&gt;Mask2Outline converts masks into FISH-Quant-compatible outlines. A Mask Editor supports manual edits or automatic mother-bud separation with a documented CNN and U-Net module (models on GitHub Releases). Optional Shift Analyzer and Registration tools correct offset between DIC-derived masks and smFISH images. A training and validation dataset of DIC images and masks is on Zenodo (&lt;a href="https://doi.org/10.5281/zenodo.17204942" rel="noopener noreferrer"&gt;doi:10.5281/zenodo.17204942&lt;/a&gt;). The paper frames the tool as accessible to researchers with minimal programming experience; a macOS installer is provided in the releases.&lt;/p&gt;

&lt;p&gt;YeastSAM is the segmentation component, not the RNA or organelle analysis, and it is not a general segmenter for non-yeast tissues.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the abstract reports
&lt;/h2&gt;

&lt;p&gt;YeastSAM achieved more than three-fold higher accuracy in segmenting dividing cells compared with existing methods. The project page does not invent additional unpublished metrics or per-dataset tables, and this post follows that boundary.&lt;/p&gt;

&lt;p&gt;A bounded claim that stays faithful to the site and the abstract is therefore:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;YeastSAM, built by adapting the Segment Anything Model (SAM) to budding yeast, achieved more than three-fold higher accuracy in segmenting dividing cells compared with existing methods, so dividing mother-bud cells are kept as correct instances rather than split into two cells.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;That sentence is intentionally scoped. It names the SAM adaptation, the organism, and the dividing-cell comparison as reported. It does not say that the same factor will appear on every microscope or stain, or that well-separated G1 cells improved by the same amount. Exact dataset splits and baselines are in the full paper; “more than three-fold” is not a guarantee on a new setup without re-evaluation.&lt;/p&gt;

&lt;p&gt;If you cite only “&amp;gt;3× more accurate,” please keep “dividing cells” and “compared with existing methods in the reported evaluation” in the same sentence.&lt;/p&gt;

&lt;p&gt;The abstract also states a downstream use, not just a segmentation score. When combined with single-molecule RNA imaging and organelle imaging, YeastSAM can be incorporated into a computational pipeline to build cellular maps, enabling quantitative analysis of the spatial regulation of gene expression. YeastSAM itself is the segmentation component; map-building and biological interpretation depend on the user’s downstream tools, imaging quality, and analysis code.&lt;/p&gt;

&lt;h2&gt;
  
  
  What has to be true for this to be relevant
&lt;/h2&gt;

&lt;p&gt;Three practical conditions sit behind the result.&lt;/p&gt;

&lt;p&gt;First, the images should be budding yeast, and the error that matters should be instance identity of dividing mother-bud pairs. If you only need a coarse foreground mask, a yeast-specific SAM adaptation is not what the paper is arguing for.&lt;/p&gt;

&lt;p&gt;Second, follow the documented install and weights. Hardware, GPU, and image-format requirements are in the docs, not in this post. Skipping the custom yeast weights is not the evaluation the paper is about.&lt;/p&gt;

&lt;p&gt;Third, if you care about smFISH or organelle channels, plan for registration. The documented pipeline generates masks from DIC and then, if needed, corrects offset before converting to FISH-Quant outlines. Segmentation accuracy on DIC does not automatically give you aligned RNA spots.&lt;/p&gt;

&lt;h2&gt;
  
  
  When not to use it
&lt;/h2&gt;

&lt;p&gt;Skip YeastSAM if your images are not budding yeast, if you only need coarse foreground masks without instance identity, or if a validated in-house model already handles mother-bud pairs on your setup. The paper’s claims focus on budding yeast dividing-cell accuracy and accessibility, not universal microscopy segmentation across tissues.&lt;/p&gt;

&lt;p&gt;Domain shift across microscopes and stains can reduce the reported “more than three-fold” gain until the method is fine-tuned or re-benchmarked on local data. Downstream cellular-map analyses depend on RNA or organelle imaging quality and on the user’s own pipeline. &lt;br&gt;
General tools — for example Cellpose-style models or vanilla SAM — are broader. The project page lists them as complementary context, not as a scored leaderboard, and this post does not invent comparison numbers for them.&lt;/p&gt;

&lt;h2&gt;
  
  
  A cautious way to use the work
&lt;/h2&gt;

&lt;p&gt;If you image budding yeast and downstream quantification fails when a mother and bud are split into two cells, the design in the paper is worth trying: run the documented YeastSAM workflow, inspect dividing cells specifically, and only then pass masks into RNA or organelle analysis. Compare against whatever you currently use on the same dividing-cell images, rather than quoting the published factor as if it had already been measured on your microscope. The implementation and docs are in the &lt;a href="https://github.com/YonghaoZhao722/YeastSAM" rel="noopener noreferrer"&gt;YeastSAM repository&lt;/a&gt; and at &lt;a href="https://yeastsamdoc.readthedocs.io/" rel="noopener noreferrer"&gt;yeastsamdoc.readthedocs.io&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;If your cells are not budding yeast, this paper does not give you a pretrained generalist. If a local model already keeps mother-bud pairs intact, it does not argue that you must switch. If you do have the yeast morphology problem, the evidence is exactly as wide as the study: a SAM adaptation optimized for budding yeast, more than three-fold higher dividing-cell accuracy versus existing methods in the reported evaluation, and a documented path into cellular maps with single-molecule RNA and organelle imaging.&lt;/p&gt;

&lt;h2&gt;
  
  
  Reporting the experiment so others can judge it
&lt;/h2&gt;

&lt;p&gt;If you benchmark a similar yeast segmenter, the comparison that matters is dividing-cell instance quality, not only whole-field scores on easy, well-separated cells. Write down the microscope, the imaging modality (the documented mask workflow is DIC), whether custom yeast weights were used, and how mother-bud correctness was scored. If you also report RNA or organelle maps, say how masks were registered to those channels.&lt;/p&gt;

&lt;p&gt;Do not collapse the paper’s claim into “YeastSAM is three times better at cell segmentation.” The published factor is about dividing cells, relative to the paper’s existing-method comparison.&lt;/p&gt;

&lt;h2&gt;
  
  
  A note on scope for practitioners
&lt;/h2&gt;

&lt;p&gt;People may want to cite this paper for a very broad sentence: foundation segmentation models can be adapted to difficult cell morphologies. That sentence is downstream of the result, but it is broader than the evidence. The evidence is: budding yeast, mother-bud errors, a SAM adaptation, more than three-fold higher accuracy on dividing cells versus existing methods, and an accessible workflow that can sit in front of single-molecule RNA and organelle imaging. If your use case matches that structure, the paper is directly on point. If it does not, treat the citation as a related-work pointer and look for evidence in your own domain.&lt;/p&gt;

&lt;p&gt;The journal version, with full details and limitations, is &lt;a href="https://doi.org/10.1091/mbc.E25-09-0454" rel="noopener noreferrer"&gt;DOI 10.1091/mbc.E25-09-0454&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>deeplearning</category>
      <category>bioinformatics</category>
      <category>python</category>
      <category>microscopy</category>
    </item>
    <item>
      <title>Single-cell survival analysis when your cohort only has bulk RNA-seq</title>
      <dc:creator>ML notes</dc:creator>
      <pubDate>Sat, 21 Nov 2026 21:19:53 +0000</pubDate>
      <link>https://dev.to/mlnotes_bio/single-cell-survival-analysis-when-your-cohort-only-has-bulk-rna-seq-4khg</link>
      <guid>https://dev.to/mlnotes_bio/single-cell-survival-analysis-when-your-cohort-only-has-bulk-rna-seq-4khg</guid>
      <description>&lt;h1&gt;
  
  
  Single-cell survival analysis when your cohort only has bulk RNA-seq
&lt;/h1&gt;

&lt;p&gt;Large cancer cohorts with survival follow-up are usually profiled by bulk RNA-seq. Single-cell RNA-seq can capture tumor heterogeneity, but it is rarely available together with patient survival outcomes. It is tempting to treat “use single-cell methods on clinical survival data” as if a patient-level scRNA-seq cohort already existed. This post stays inside one published study and asks a narrower question: when the cohort only has bulk RNA-seq, what does reconstructing cellular profiles actually improve?&lt;/p&gt;

&lt;p&gt;The paper is Zhao et al., &lt;em&gt;DeSCENT: Deconvolutional Single-Cell RNA-seq Enhances Transcriptome-based Cancer Survival Analysis&lt;/em&gt;, &lt;em&gt;Bioinformatics&lt;/em&gt;, 2026, article btag810 (&lt;a href="https://doi.org/10.1093/bioinformatics/btag810" rel="noopener noreferrer"&gt;DOI 10.1093/bioinformatics/btag810&lt;/a&gt;). The code is at &lt;a href="https://github.com/YonghaoZhao722/DeSCENT" rel="noopener noreferrer"&gt;https://github.com/YonghaoZhao722/DeSCENT&lt;/a&gt;. An archival snapshot is at &lt;a href="https://doi.org/10.5281/zenodo.22073099" rel="noopener noreferrer"&gt;doi:10.5281/zenodo.22073099&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  The setting, not a slogan
&lt;/h2&gt;

&lt;p&gt;DeSCENT is a bulk-to-single-cell framework for cancer survival analysis. It reconstructs reference-informed cellular expression profiles from bulk RNA-seq, then combines the original bulk representation with those reconstructed cellular representations for multimodal survival prediction.&lt;/p&gt;

&lt;p&gt;Two inputs are required. The target cohort needs bulk RNA-seq plus survival labels — the familiar TCGA-style setting. It also needs a same-cancer annotated scRNA-seq reference, used for deconvolution and generation. Without a matched-cancer reference atlas, the method is not specified to run as described. The public repository ships a BRCA demo; verified cancer configurations in the repository are BRCA, COAD, HNSC, KIRC, LGG, LIHC, LUAD, and STAD. Demo bulk and single-cell assets are distributed via Zenodo rather than Git, because of file size.&lt;/p&gt;

&lt;p&gt;This is not a claim that reconstructed profiles are measured patient scRNA-seq, and it is not a claim that every deconvolution tool is thereby a survival model.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the pipeline actually does
&lt;/h2&gt;

&lt;p&gt;The open pipeline has four stages.&lt;/p&gt;

&lt;p&gt;First, ReDeconv estimates cell fractions from each bulk sample together with the same-cancer reference. The repository bundles a patched ReDeconv fork for this step; the project page treats it as the deconvolution stage of the published pipeline, not as a standalone replacement for every cell-fraction method.&lt;/p&gt;

&lt;p&gt;Second and third, scDiffusion is trained (or loaded from checkpoints) and then used for conditional generation of patient-level synthetic single-cell gene-expression profiles (scGEP, stored as &lt;code&gt;.npz&lt;/code&gt; in the open code). These profiles are reference-informed estimates.&lt;/p&gt;

&lt;p&gt;Fourth, multimodal survival models are trained on bulk plus reconstructed cellular features and evaluated with cross-validation. The paper reports C-index and related metrics. The repository documents a multimodal 5-fold CV entry point and a pure-bulk baseline; users must supply paths and pretrained assets as the README describes. Entry points include a BRCA notebook demo and the shell scripts &lt;code&gt;run_part1_deg_redeconv_condgen.sh&lt;/code&gt; and &lt;code&gt;run_part2_survival.sh&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;If you only need a one-sentence description of DeSCENT: it is a way to bring reconstructed, reference-informed cellular profiles into a bulk-cohort survival model. It is not a substitute for a real single-cell clinical cohort when one already exists.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the abstract reports
&lt;/h2&gt;

&lt;p&gt;The &lt;em&gt;Bioinformatics&lt;/em&gt; abstract evaluates DeSCENT with three survival models on eight TCGA cancer cohorts. Compared with Bulk, DeSCENT improved average C-index and generally improved time-dependent AUC, while integrated Brier score results were mixed. External validation showed higher average C-index than all three baselines, although performance was lower than in internal evaluation. The abstract’s own conclusion is modest: these findings support the use of cellular representations in the evaluated transcriptome-only setting.&lt;/p&gt;

&lt;p&gt;The project page warns against extrapolating exact per-cancer C-index tables unless the full paper is read. This post follows that boundary and does not invent point values that the abstract does not report.&lt;/p&gt;

&lt;p&gt;A bounded claim that stays faithful to the site and the abstract is therefore:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;DeSCENT reconstructs reference-informed cellular expression profiles from bulk RNA-seq and combines bulk and reconstructed cellular representations for multimodal survival prediction, improving average C-index over bulk-only baselines across eight TCGA cancer cohorts (time-dependent AUC generally improved; integrated Brier score results were mixed).&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;That sentence is intentionally scoped. It names the reconstruction step, the fusion with bulk, the eight TCGA cohorts, and the mixed IBS result. It does not say that every cancer, every external cohort, or every calibration-sensitive use case saw a uniform gain. It does not treat reconstructed profiles as ground-truth single-cell expression.&lt;/p&gt;

&lt;p&gt;If you cite only “DeSCENT improves C-index,” please keep the Bulk comparison, the eight-cohort setting, and the mixed integrated Brier score in the same sentence.&lt;/p&gt;

&lt;h2&gt;
  
  
  What has to be true for this to be relevant
&lt;/h2&gt;

&lt;p&gt;Three practical conditions sit behind the result.&lt;/p&gt;

&lt;p&gt;First, you need bulk RNA-seq with survival labels. DeSCENT is aimed at TCGA-style cohorts that already carry follow-up. A reference atlas alone is not a survival study.&lt;/p&gt;

&lt;p&gt;Second, the scRNA-seq reference has to be the same cancer type as the bulk cohort. The method as described is not a license to borrow an unrelated tissue atlas and call the output patient-level single-cell data.&lt;/p&gt;

&lt;p&gt;Third, the reconstructed profiles have to be treated as what they are: reference-informed estimates used as an additional representation in a survival model. They do not replace measured patient-level scRNA-seq.&lt;/p&gt;

&lt;h2&gt;
  
  
  When not to use it
&lt;/h2&gt;

&lt;p&gt;Do not use DeSCENT as a substitute for measured patient-level scRNA-seq when such survival cohorts already exist — prefer methods built for that setting, such as scSurvival / scSurv-style models. Do not use it if you only need cell-type fractions or phenotype-associated subpopulations and do not need a survival-prediction model; BayesPrism / CIBERSORTx and Scissor / DEGAS are aimed at those tasks, and the project page does not invent head-to-head benchmark numbers against them. Do not use it without a matched-cancer reference, or if you need ground-truth single-cell expression rather than reconstructed profiles.&lt;/p&gt;

&lt;p&gt;Integrated Brier score gains were mixed in the paper, so calibration-sensitive use cases need careful checking. External validation beat three baselines on average C-index but was lower than internal evaluation; that drop should be expected and reported, not hidden.&lt;/p&gt;

&lt;h2&gt;
  
  
  A cautious way to use the work
&lt;/h2&gt;

&lt;p&gt;If you have a bulk RNA-seq cohort with survival labels and a same-cancer annotated scRNA-seq reference, the design in the paper is worth trying: reconstruct reference-informed cellular profiles, fuse them with the original bulk representation, and compare against a bulk-only model on the same patients and the same metrics. Report C-index, time-dependent AUC, and integrated Brier score, not only the metric that improved. The implementation is in the &lt;a href="https://github.com/YonghaoZhao722/DeSCENT" rel="noopener noreferrer"&gt;DeSCENT repository&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;If you already have a large patient-level scRNA-seq survival cohort, this paper does not argue that you should discard it in favor of reconstruction. If you do not have a same-cancer reference, it does not give you a workaround. If you do have both inputs, the evidence is exactly as wide as the study: eight TCGA cohorts, three survival models, improved average C-index versus Bulk, generally improved time-dependent AUC, mixed IBS, and a weaker but still higher-than-baseline average C-index on external validation.&lt;/p&gt;

&lt;h2&gt;
  
  
  Reporting the experiment so others can judge it
&lt;/h2&gt;

&lt;p&gt;If you run a similar bulk-to-single-cell survival analysis, the comparison that matters is the same survival model with and without the reconstructed cellular representation, on the same patients and the same metrics. Write down the bulk cohort, the reference atlas, how the two were matched by cancer type, and whether evaluation was internal cross-validation or an external cohort. Because IBS was mixed in the published study, omitting it and reporting only C-index would hide part of the result.&lt;/p&gt;

&lt;p&gt;The abstract does not name the three survival models or the three external-validation baselines, and it does not publish a single headline C-index. Quote the directional findings that it does report, and point readers to the paper for per-cohort tables. The open repository includes reproduction commands and a pure-bulk baseline; if you use those scripts, say so, and keep head-to-head comparisons matched as the README requires.&lt;/p&gt;

&lt;h2&gt;
  
  
  A note on scope for practitioners
&lt;/h2&gt;

&lt;p&gt;People may want to cite this paper for a very broad sentence: single-cell information improves transcriptome-based survival models. That sentence is downstream of the result, but it is broader than the evidence. The evidence is: reference-informed reconstructed cellular profiles, fused with bulk, on eight TCGA cohorts, with improved average C-index versus Bulk, generally improved time-dependent AUC, mixed integrated Brier score, and external validation that was better than three baselines but weaker than the internal evaluation. If your use case matches that structure, the paper is directly on point. If it does not, treat the citation as a related-work pointer and look for evidence in your own domain.&lt;/p&gt;

&lt;p&gt;The journal version, with full details and limitations, is &lt;a href="https://doi.org/10.1093/bioinformatics/btag810" rel="noopener noreferrer"&gt;DOI 10.1093/bioinformatics/btag810&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>bioinformatics</category>
      <category>machinelearning</category>
      <category>genomics</category>
      <category>datascience</category>
    </item>
    <item>
      <title>Transfer learning for small-sample survival data: what works at n=50</title>
      <dc:creator>ML notes</dc:creator>
      <pubDate>Sat, 21 Nov 2026 21:12:25 +0000</pubDate>
      <link>https://dev.to/mlnotes_bio/transfer-learning-for-small-sample-survival-data-what-works-at-n50-4jhb</link>
      <guid>https://dev.to/mlnotes_bio/transfer-learning-for-small-sample-survival-data-what-works-at-n50-4jhb</guid>
      <description>&lt;h1&gt;
  
  
  Transfer learning for small-sample survival data: what works at n=50
&lt;/h1&gt;

&lt;p&gt;Hospital cancer cohorts are often too small to fit a stable survival model. Large registries exist, but they are not the same population as the patients in front of you. It is tempting to treat “transfer learning helps small-sample problems” as a general fact. This post stays inside one published study and asks a narrower question: in a related-source setting, what actually improved, and at what sample size?&lt;/p&gt;

&lt;p&gt;The paper is Zhao et al., &lt;em&gt;Tackling small sample survival analysis via transfer learning: A study of colorectal cancer prognosis&lt;/em&gt;, &lt;em&gt;Artificial Intelligence in Medicine&lt;/em&gt; 178:103426, 2026 (&lt;a href="https://doi.org/10.1016/j.artmed.2026.103426" rel="noopener noreferrer"&gt;DOI 10.1016/j.artmed.2026.103426&lt;/a&gt;; &lt;a href="https://arxiv.org/abs/2501.12421" rel="noopener noreferrer"&gt;arXiv:2501.12421&lt;/a&gt;). The code is at &lt;a href="https://github.com/YonghaoZhao722/TSF" rel="noopener noreferrer"&gt;https://github.com/YonghaoZhao722/TSF&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  The setting, not a slogan
&lt;/h2&gt;

&lt;p&gt;The source cohort is 27,379 SEER colorectal cancer (CRC) stage I patients. The target cohort is 728 CRC stage I patients from West China Hospital. Both sides are the same disease and the same stage. That relatedness is part of the result, not a minor footnote. The paper does not claim that an arbitrary large dataset will help an arbitrary small one.&lt;/p&gt;

&lt;p&gt;The clinical problem is familiar: a hospital series can be too small to induce useful survival patterns on its own, while a registry is large enough to learn structure that might transfer if the target is related. Transfer learning is used here as a way to bring source-task knowledge into the target analysis, then adapt it with the hospital data.&lt;/p&gt;

&lt;h2&gt;
  
  
  Two transfer mechanisms
&lt;/h2&gt;

&lt;p&gt;The paper studies transfer for several standard survival models, not only one architecture.&lt;/p&gt;

&lt;p&gt;For parametric and neural models — DeepSurv, Cox-CC (Cox-based neural networks), and DeepHit — the transfer recipe is the usual one: pretrain on the source cohort and fine-tune on the target. That is ordinary transfer learning applied to survival losses rather than a new forest algorithm.&lt;/p&gt;

&lt;p&gt;For Random Survival Forest (RSF), the paper proposes Transfer Survival Forest (TSF). TSF transfers tree structures learned on the source task and fine-tunes those trees with the target data. The point of TSF is that a non-parametric forest does not have a single weight vector you can pretrain in the neural-network sense, so the transferable object is the tree structure itself.&lt;/p&gt;

&lt;p&gt;If you only need a one-sentence description of TSF: it is a transfer-learning version of RSF that moves source trees to the target and then fine-tunes them. It is not a method for single-cell or bulk RNA-seq survival (that is a different paper), and it is not a substitute for a related source cohort.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the Ctd numbers say
&lt;/h2&gt;

&lt;p&gt;On the full hospital target cohort, transfer learning improved time-dependent concordance (Ctd) for every model the paper reports in the abstract:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Cox-CC: 0.7868 → 0.8111&lt;/li&gt;
&lt;li&gt;DeepHit: 0.8085 → 0.8135&lt;/li&gt;
&lt;li&gt;DeepSurv: 0.7722 → 0.8043&lt;/li&gt;
&lt;li&gt;RSF / TSF: 0.7940 → 0.8297 (the highest of these)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Those are the published point values for this CRC stage I SEER-to-hospital setup. They should not be quoted as expected gains for other diseases, stages, or feature sets. DeepHit’s lift is small; TSF’s is the largest in this table. The paper’s conclusion is not “any model plus transfer always wins by a large margin.” It is that these common survival models can be enhanced by transfer designs that fit the model class.&lt;/p&gt;

&lt;h2&gt;
  
  
  The n = 50 result
&lt;/h2&gt;

&lt;p&gt;The question in the title is about very small targets. The abstract states that all of these models, when trained with data as small as 50, demonstrated even more significant improvement under transfer learning. A bounded claim that stays faithful to the abstract is therefore:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Transfer learning from a related large cohort can improve survival models when the target clinical sample is small (gains observed down to n = 50 in a SEER-to-hospital colorectal cancer study).&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;That sentence is intentionally scoped. It names the sample size, the relatedness assumption, and the study design. It does not say that n = 50 is enough for a high-dimensional genomic model, that the improvement will have the same size as on the n = 728 cohort, or that TSF is the only method that worked. Neural models in the same paper also improved with pretraining and fine-tuning.&lt;/p&gt;

&lt;p&gt;If you cite only “n = 50 works,” please keep the related-source condition in the same sentence.&lt;/p&gt;

&lt;h2&gt;
  
  
  What has to be true for this to be relevant
&lt;/h2&gt;

&lt;p&gt;Three practical conditions sit behind the result.&lt;/p&gt;

&lt;p&gt;First, the source and target have to be related. Here both are CRC stage I. Negative transfer is possible if the domains mismatch. A large but distant source — a different cancer, a different outcome, or a very different covariate distribution — is outside the evidence of this study.&lt;/p&gt;

&lt;p&gt;Second, features have to be harmonized. Registry fields and hospital fields are rarely identical. The paper’s limitations treat feature harmonization as a practical bottleneck that is not solved merely by naming TSF or turning on fine-tuning. If you cannot align a usable feature set, you do not yet have the experimental setting in which these Ctd gains were observed.&lt;/p&gt;

&lt;p&gt;Third, use the transfer mechanism that matches the model. TSF is the RSF path. DeepSurv, Cox-CC, and DeepHit use pretraining and fine-tuning. Swapping those recipes without looking at the model class is not what the paper did.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this study is not
&lt;/h2&gt;

&lt;p&gt;It is not a proof that transfer learning always beats a well-regularized model trained only on the target. It is not a license to ignore events-per-variable or validation. Concordance on one transfer pair is not a clinical deployment claim. And it is not advice to scrape an unrelated public dataset and call it pretraining.&lt;/p&gt;

&lt;p&gt;It is also not an argument against simple models when you have very few events. The n = 50 experiments still sit inside a related, already-defined CRC stage I transfer. If you have 30 patients, a panel of many variants, and no aligned registry source, the first problem is usually feature collapse and honest uncertainty, not which forest to fine-tune.&lt;/p&gt;

&lt;h2&gt;
  
  
  A cautious way to use the work
&lt;/h2&gt;

&lt;p&gt;If you have a small hospital survival cohort and a large, related registry with overlapping covariates, the design in the paper is worth trying: pretrain or transfer structure on the source, fine-tune on the target, and compare against the same model trained on the target alone. Report the target-only baseline. Keep the relatedness of the source explicit. If you use RSF, TSF is the method introduced for that case; the implementation is in the &lt;a href="https://github.com/YonghaoZhao722/TSF" rel="noopener noreferrer"&gt;TSF repository&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;If you do not have a related source, this paper does not give you a workaround. If you do have one, the evidence that transfer can still help at n = 50 is exactly as wide as the study: CRC stage I, SEER to one hospital cohort, the models and metrics above.&lt;/p&gt;

&lt;h2&gt;
  
  
  Reporting the experiment so others can judge it
&lt;/h2&gt;

&lt;p&gt;If you run a similar transfer, the comparison that matters is the &lt;em&gt;same&lt;/em&gt; survival model with and without transfer, on the same target patients and the same metric. The abstract of this paper reports Ctd, not a generic claim about every concordance estimator. Write down the source cohort, the inclusion rules (here, CRC stage I on both sides), how features were aligned, and the target sample size used for each run. When you subsample the target — as this study did down to n = 50 — say how the subset was formed and keep the source fixed. Otherwise a reader cannot tell whether the extra lift at small n is a property of transfer or a property of a lucky draw.&lt;/p&gt;

&lt;p&gt;The target-only baseline should still be reported even if transfer is expected to help. The DeepHit gain in this study is modest (0.8085 to 0.8135). Publishing only the transferred number would have hidden that. TSF reaching 0.8297 is the top of &lt;em&gt;this&lt;/em&gt; table, not a guarantee that a forest will dominate a neural model on the next data set.&lt;/p&gt;

&lt;h2&gt;
  
  
  A note on scope for practitioners
&lt;/h2&gt;

&lt;p&gt;People sometimes cite this paper to support a very broad sentence: transfer learning is an effective strategy for small-sample learning. That sentence is downstream of the result, but it is broader than the evidence. The evidence is: related source, CRC stage I, SEER to one hospital cohort, four model families, Ctd gains, and a stronger relative improvement when the target was reduced to as few as 50 samples. If your use case matches that structure, the paper is directly on point. If it does not, treat the citation as a related-work pointer and look for evidence in your own domain.&lt;/p&gt;

&lt;p&gt;The journal version, with full details and limitations, is &lt;a href="https://doi.org/10.1016/j.artmed.2026.103426" rel="noopener noreferrer"&gt;DOI 10.1016/j.artmed.2026.103426&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>machinelearning</category>
      <category>survivalanalysis</category>
      <category>healthcare</category>
      <category>datascience</category>
    </item>
  </channel>
</rss>
