MB-DSMIL-CL-PL: Scalable Weakly Supervised Ovarian Cancer Subtype Classification and Localisation Using Contrastive and Prototype Learning with Frozen Patch Features

Marcus Jenkins; Jasenka Mazibrada; Bogdan Leahu; Michal Mackiewicz

cs.CVFeb 16, 2026

MB-DSMIL-CL-PL: Scalable Weakly Supervised Ovarian Cancer Subtype Classification and Localisation Using Contrastive and Prototype Learning with Frozen Patch Features

Authors: Marcus Jenkins, Jasenka Mazibrada, Bogdan Leahu, Michal Mackiewicz

Organizations: School of Computing Sciences, University of East Anglia, Norwich, Norfolk, NR4 7TJ, UK · Norfolk and Norwich University Hospital, Norwich, Norfolk, NR4 7UY, UK

Abstract

The study of histopathological subtypes is valuable for the personalisation of effective treatment strategies for ovarian cancer. However, increasing diagnostic workloads present a challenge for UK pathology departments, leading to the rise in AI approaches. While traditional approaches in this field have relied on pre-computed, frozen image features, recent advances have shifted towards end-to-end feature extraction, providing an improvement in accuracy but at the expense of significantly reduced scalability during training and time-consuming experimentation. In this paper, we propose a new approach for subtype classification and localisation in ovarian cancer histopathology images using contrastive and prototype learning with pre-computed, frozen features via feature-space augmentations. Compared to DSMIL, our method achieves an improvement of 70.4% and 15.3% in F1 score for instance- and slide-level classification, respectively, along with AUC gains of 16.9% for instance localisation and 2.3% for slide classification, while maintaining the use of frozen patch features.

Figures & tables

Figure 1: Architecture of the MB-DSMIL-CL contrastive learning. Instance classifier predictions are not used for the MoCo warmup epochs

Figure 2: Class prototype learning with precomputed instance features in MB-DSMIL-CL-PL.

Split	HGSC	LGSC	CCC	EA	MBT	SBT	MA	Normal	Total
Train	35	10	6	8	15	11	4	11	100
Test	11	4	3	4	3	5	2	5	37
Total	46	14	9	12	18	16	6	16	137

Table 1: Slide-level label distribution for the training and test sets.

Model	Slide (Macro F1)	Instance (Macro F1)	GT vs. Normal (Macro F1)
DSMIL	0.672 $\pm$ 0.0926	0.301 $\pm$ 0.0329	0.567 $\pm$ 0.0584
CLAM	0.757 $\pm$ 0.0703	0.309 $\pm$ 0.0701	0.376 $\pm$ 0.0784
MB-DSMIL-CL-PL	0.775 $\pm$ 0.0469	0.513 $\pm$ 0.0214	0.684 $\pm$ 0.0297
DSMIL-CL-PL*	0.698 $\pm$ 0.0553	0.483 $\pm$ 0.0101	0.655 $\pm$ 0.0273

Table 2: Comparison of DSMIL, CLAM, and MB-DSMIL-CL-PL (ours) on the slide- and instance-level classification tasks using the mean and standard deviation (mean

\pm

std) of macro F1 for each model from each fold on the test set. GT vs. Noraml indicates that only the slide GT label and normal tissue class logits are considered for the instance classifier’s predictions. Bold denotes the best value in each column and italics indicates the second-best. Ablation experiments are denoted by *.

Model	Slide (OvR AUC)	Instance (OvR AUC)	GT vs. Normal (AUC)	Attention (AUC)
DSMIL	0.952 $\pm$ 0.0196	0.781 $\pm$ 0.0373	0.580 $\pm$ 0.0534	0.756 $\pm$ 0.0405
MB-DSMIL-CL-PL	0.974 $\pm$ 0.00689	0.913 $\pm$ 0.00981	0.807 $\pm$ 0.0242	0.876 $\pm$ 0.00760
DSMIL-CL-PL	0.949 $\pm$ 0.00988	0.873 $\pm$ 0.0103	0.782 $\pm$ 0.0171	0.854 $\pm$ 0.0126

Table 3: Comparison of DSMIL, CLAM, and MB-DSMIL-CL-PL (ours) on the slide-classification and localisation tasks using the mean and standard deviation (mean

\pm

std) of ROC AUC for each model from each fold on the test set. Bold denotes the best value in each column and italics denotes the second-best.

Class	Slide F1		Instance F1
	DSMIL	MB-DSMIL-CL-PL	DSMIL	MB-DSMIL-CL-PL
MucA	0.327 $\pm$ 0.306	0.733 $\pm$ 0.389	0.183 $\pm$ 0.037	0.417 $\pm$ 0.092
CCC	0.840 $\pm$ 0.080	0.620 $\pm$ 0.147	0.435 $\pm$ 0.067	0.636 $\pm$ 0.102
EA	0.455 $\pm$ 0.376	0.740 $\pm$ 0.111	0.174 $\pm$ 0.046	0.392 $\pm$ 0.070
LGSC	0.583 $\pm$ 0.149	0.590 $\pm$ 0.067	0.276 $\pm$ 0.081	0.400 $\pm$ 0.040
SBT	0.777 $\pm$ 0.077	0.806 $\pm$ 0.070	0.132 $\pm$ 0.054	0.288 $\pm$ 0.018

Table 4: Per-class slide- and instance-level F1 scores on the test set for the model from each fold (mean

\pm

std). Best performance per class and task is shown in bold.

Panel of Figure 3 — Figure 3: Qualitative comparison for an MBT slide (test set). The ground-truth image is identical across methods and shown in the left column for reference. Rows show instance predictions, predictions filtered by GT vs. Normal, and attention maps (scaled using Min-Max normalisation).

Panel of Figure 4 — Figure 4: Qualitative comparison for a HGSC slide (test set). The ground-truth image is identical across methods and shown in the left column for reference.

Panel of Figure 5 — Figure 5: Qualitative comparison for an SBT slide (test set). The ground-truth image is identical across methods and shown in the left column for reference.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Value
Learning rate	0.0001
Weight decay	0.001
Query embedding dim ( $q$ )	64
Feature size ( $e$ )	1024
$\eta_{w}$	0.05
$\eta_{s}$	0.4

Appendix

Table 5: Optimal hyperparameter configuration for MB-DSMIL-CL-PL for our dataset.

Panel of Figure 6 — Figure 6: Normalised confusion matrices for slide classifications for (a) DSMIL and (b) MB-DSMIL-CL-PL, for instance classifications for (c) DSMIL and (d) MB-DSMIL-CL-PL, and for filtered instance classifications for (e) DSMIL and (f) MB-DSMIL-CL-PL.

Histopathology Image Classification

Histopathology image classification uses computational models to identify tissue types, tumors, grades, and biomarkers from stained pathology images.

See all 139 papers

PC-MIL: Decoupling Feature Resolution from Supervision Scale in Whole-Slide Learning
Apr 13, 2026Syed Fahim Ahmed, Gnanesh Rasineni, Florian Koehler +7Whole-Slide Image Analysis Histopathology Image Classification
Kahlert School of Computing, University of Utah, Salt Lake City, UT 84112, USA · Scientific Computing & Imaging Institute, University of Utah, Salt Lake City, UT 84112, USA · Instapath Inc., Houston, TX 77021, USA +1
Less Contouring, More Accuracy: Lesion-Guided ROI Deep Learning for Ovarian Ultrasound Classification
Aug 26, 2026Mehran Ahmad, Ali Abbasian Ardakani, Afshin Mohammadi +3Medical Image Classification Ultrasound Imaging
Department of Medicine, Danube Private University (DPU), Krems, Austria. · Austrian Centre for Medical Innovation and Technology (ACMIT), Wiener Neustadt, Austria. · Department of Medical Physics and Biomedical Engineering, Medical University of Vienna, Vienna, Austria. +1
Automated Detection of Malignant Lesions in the Ovary Using Deep Learning Models and XAI
Mar 12, 2026Md. Hasin Sarwar Ifty, Nisharga Nirjan, Labib Islam +4Histopathology Image Classification Computational Pathology
Dept. of Computer Science and Engr. BRAC University Dhaka, Bangladesh

Apr 13, 2026J/K move · Enter open · S save

PC-MIL: Decoupling Feature Resolution from Supervision Scale in Whole-Slide Learning

Syed Fahim Ahmed, Gnanesh Rasineni, Florian Koehler, Abu Zahid Bin Aziz+6

Kahlert School of Computing, University of Utah, Salt Lake City, UT 84112, USA · Scientific Computing & Imaging Institute, University of Utah, Salt Lake City, UT 84112, USA · Instapath Inc., Houston, TX 77021, USA+1

Whole-Slide Image Analysis Histopathology Image Classification Multiple Instance Learning Weakly Supervised Learning

Open paper AskPDF arXiv

Whole-slide image (WSI) classification in computational pathology is commonly formulated as slide-level Multiple Instance Learning (MIL) with a single global bag representation. However, slide-level MIL is fundamentally underconstrained: optimizing only global labels encourages models to aggregate features without learning anatomically meaningful localization. This creates a mismatch between the scale of supervision and the scale of clinical reasoning. Clinicians assess tumor burden, focal lesions, and architectural patterns within millimeter-scale regions, whereas standard MIL is trained only to predict whether "somewhere in the slide there is cancer." As a result, the model's inductive bias effectively erases anatomical structure. We propose Progressive-Context MIL (PC-MIL), a framework that treats the spatial extent of supervision as a first-class design dimension. Rather than altering magnification, patch size, or introducing pixel-level segmentation, we decouple feature resolution from supervision scale. Using fixed 20x features, we vary MIL bag extent in millimeter units and anchor supervision at a clinically motivated 2mm scale to preserve comparable tumor burden and avoid confounding scale with lesion density. PC-MIL progressively mixes slide- and region-level supervision in controlled proportions, enabling explicit train-context x test-context analysis. On 1,476 prostate WSIs from five public datasets for binary cancer detection, we show that anatomical context is an independent axis of generalization in MIL, orthogonal to feature resolution: modest regional supervision improves cross-context performance, and balanced multi-context training stabilizes accuracy across slide and regional evaluation without sacrificing global performance. These results demonstrate that supervision extent shapes MIL inductive bias and support anatomically grounded WSI generalization.

Abstract

Histopathology Image Classification

Explore similar work

PC-MIL: Decoupling Feature Resolution from Supervision Scale in Whole-Slide Learning

Less Contouring, More Accuracy: Lesion-Guided ROI Deep Learning for Ovarian Ultrasound Classification

Automated Detection of Malignant Lesions in the Ovary Using Deep Learning Models and XAI