Interpretable deep learning model to predict the molecular classification of endometrial cancer from haematoxylin and eosin-stained whole-slide images: a combined analysis of the PORTEC randomised trials and clinical cohorts
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
BackgroundEndometrial cancer can be molecularly classified into POLEmut, mismatch repair deficient (MMRd), p53 abnormal (p53abn), and no specific molecular profile (NSMP) subgroups. We aimed to develop an interpretable deep learning pipeline for whole-slide-image-based prediction of the four molecular classes in endometrial cancer (im4MEC), to identify morpho-molecular correlates, and to refine prognostication.MethodsThis combined analysis included diagnostic haematoxylin and eosin-stained slides and molecular and clinicopathological data from 2028 patients with intermediate-to-high-risk endometrial cancer from the PORTEC-1 (n=466), PORTEC-2 (n=375), and PORTEC-3 (n=393) randomised trials and the TransPORTEC pilot study (n=110), the Medisch Spectrum Twente cohort (n=242), a case series of patients with POLEmut endometrial cancer in the Leiden Endometrial Cancer Repository (n=47), and The Cancer Genome Atlas-Uterine Corpus Endometrial Carcinoma cohort (n=395). PORTEC-3 was held out as an independent test set and a four-fold cross validation was performed. Performance was measured with the macro and class-wise area under the receiver operating characteristic curve (AUROC). Whole-slide images were segmented into tiles of 360 μm resized to 224 × 224 pixels. im4MEC was trained to learn tile-level morphological features with self-supervised learning and to molecularly classify whole-slide images with an attention mechanism. The top 20 tiles with the highest attention scores were reviewed to identify morpho-molecular correlates. Predictions of a nuclear classification deep learning model serve to derive interpretable morphological features. We analysed 5-year recurrence-free survival and explored prognostic refinement by molecular class using the Kaplan-Meier method.Findingsim4MEC attained macro-average AUROCs of 0·874 (95% CI 0·856–0·893) on four-fold cross-validation and 0·876 on the independent test set. The class-wise AUROCs were 0·849 for POLEmut (n=51), 0·844 for MMRd (n=134), 0·883 for NSMP (n=120), and 0·928 for p53abn (n=88). POLEmut and MMRd tiles had a high density of lymphocytes, p53abn tiles had strong nuclear atypia, and the morphology of POLEmut and MMRd endometrial cancer overlapped. im4MEC highlighted a low tumour-to-stroma ratio as a potentially novel characteristic feature of the NSMP class. 5-year recurrence-free survival was significantly different between im4MEC predicted molecular classes in PORTEC-3 (log-rank p<0·0001). The ten patients with aggressive p53abn endometrial cancer that was predicted as MMRd showed inflammatory morphology and appeared to have a better prognosis than patients with correctly predicted p53abn endometrial cancer (p=0·30). The four patients with NSMP endometrial cancer that was predicted as p53abn showed higher nuclear atypia and appeared to have a worse prognosis than patients with correctly predicted NSMP (p=0·13). Patients with MMRd endometrial cancer predicted as POLEmut had an excellent prognosis, as do those with true POLEmut endometrial cancer.InterpretationWe present the first interpretable deep learning model, im4MEC, for haematoxylin and eosin-based prediction of molecular endometrial cancer classification. im4MEC robustly identified morpho-molecular correlates and could enable further prognostic refinement of patients with endometrial cancer.FundingThe Hanarth Foundation, the Promedica Foundation, and the Swiss Federal Institutes of Technology.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.003 | 0.003 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.002 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it