Leveraging whole slide difficulty in Multiple Instance Learning to improve prostate cancer grading

Marie Arrivat; Rémy Peyret; Elsa Angelini; Pietro Gori

cs.CVMar 10, 2026

Leveraging whole slide difficulty in Multiple Instance Learning to improve prostate cancer grading

Authors: Marie Arrivat, Rémy Peyret, Elsa Angelini, Pietro Gori

Organizations: Primaa, Paris, France · LTCI, Telecom Paris, Institut Polytechnique de Paris, France

Abstract

Multiple Instance Learning (MIL) has been widely applied in histopathology to classify Whole Slide Images (WSIs) with slide-level diagnoses. While the ground truth is established by expert pathologists, the slides can be difficult to diagnose for non-experts and lead to disagreements between the annotators. In this paper, we introduce the notion of Whole Slide Difficulty (WSD), based on the disagreement between an expert and a non-expert pathologist. We propose two different methods to leverage WSD, a multi-task approach and a weighted classification loss approach, and we apply them to Gleason grading of prostate cancer slides. Results show that integrating WSD during training consistently improves the classification performance across different feature encoders and MIL methods, particularly for higher Gleason grades (i.e. worse diagnosis).

Figures & tables

	MaxMIL		ABMIL		CLAM		DSMIL		TransMIL
Methods	Bal. Acc.	W. F1-Score	Bal. Acc.	W. F1-Score	Bal. Acc.	W. F1-Score	Bal. Acc.	W. F1-Score	Bal. Acc.	W. F1-Score
Classification baseline	68.3	75.4	71.5	79.7	72.1	79.8	70.4	78.4	73.0	81.6
Multi-task learning	67.8 (-0.7, +0.7)	75.2 (-0.7, +0.6)	73.2 (-1.5, +1.4)	80.1 (-1.7, +1.5)	70.2 (-2.4, +2.4)	78.7 (-2.3, +2.5)	72.1 (-2.5, +2.3)	79.7 (-2.6, +2.3)	70.5 (-1.0, +0.8)	81.7 (-1.1, +0.8)
Weighted classification loss	68.5 (-1.7, +1.4)	75.0 (-1.8, +1.3)	75.6 (-1.4, +1.9) $*$	81.6 (-1.4, +2.1)	74.6 (-1.9, +1.9) $*$	80.0 (-1.9, +2.0) $*$	69.7 (-2.0, +1.8)	78.1 (-2.0, +1.8)	75.8 (-1.1, +1.4) $*$	82.1 (-1.0, +1.6) $*$

Table 1: Results of different methods using CTransPath as feature extractor and five MIL backbones. Values are mean over two random seeds (95% CI), where

*

indicates p-value

<

0.05. Improvements compared to the baseline are highlighted in bold.

	MaxMIL		ABMIL		CLAM		DSMIL		TransMIL
Methods	Bal. Acc.	W. F1-Score	Bal. Acc.	W. F1-Score	Bal. Acc.	W. F1-Score	Bal. Acc.	W. F1-Score	Bal. Acc.	W. F1-Score
Classification baseline	65.7	71.7	67.3	77.8	70.0	80.3	67.2	78.9	67.4	78.5
Multi-task learning	66.5 (-1.0, +1.1)	72.2 (-1.0, +1.0)	69.2 (-1.3, +1.5) $*$	78.8 (-1.4, +1.7)	70.6 (-1.4, +1.2)	80.5 (-1.8, +1.2)	70.4 (-1.5, +1.6)	79.5 (-1.5, +1.8) $*$	67.3 (-2.0, +1.9)	79.1 (-2.2, +2.0)
Weighted classification loss	66.7 (-1.8, +1.5)	73.9 (-1.8, +1.5)	67.4 (-2.6, +2.9)	75.6 (-2.9, +3.2)	69.7 (-1.3, +1.5)	79.3 (-1.3, +1.7)	69.7 (-1.8, +2.3)	77.1 (-1.8, +2.6)	69.4 (-1.4, +1.5) $*$	79.0 (-1.3, +1.6)

Table 2: Results of different methods using UNI2-h as feature extractor and five MIL backbones. Values are mean over two random seeds (95% CI), where

*

indicates p-value

<

0.05. Improvements compared to the baseline are highlighted in bold.

	MaxMIL		ABMIL		CLAM		DSMIL		TransMIL
Class	Baseline	WSD-B method	Baseline	WSD-B method	Baseline	WSD-B method	Baseline	WSD-B method	Baseline	WSD-B method
Benign	79.3	78.2	91.7	90.2	90.9	90.0	89.9	92.0	91.5	90.0
Gleason 3	56.3	53.1	58.3	65.6	62.0	62.0	59.9	56.3	63.0	68.8
Gleason 4	88.3	88.3	83.5	83.5	80.1	80.6	79.2	85.0	87.4	84.0
Gleason 5	47.4	52.6	52.6	63.2	55.3	65.8	52.6	55.3	50.0	60.5

Table 3: Accuracy per class for the classification baseline and the best WSD-based (WSD-B) method for each MIL backbone, using CTransPath as feature extractor. The best results are highlighted in bold.

Figure 1: Attention maps of ABMIL on CTransPath with the baseline method and with the weighted classification loss. This slide was labeled as 3+3 and had no consensus. The baseline model focuses on irrelevant patches and classifies the slide as Benign while the WSD-based model focuses on a patch containing a Gleason 3 gland (circled in black) and correctly classifies the slide as Gleason 3.

( $\alpha$ , $\beta$ )	(1,0)	(1,1)	(1,2)	(1,3)	(1,10)	(1,50)
Bal. Acc.	71.5	72.9	72.6	71.7	73.2	73.7
W. F1-Score	79.7	80.0	79.7	79.9	80.1	79.5

Table 4: Impact of (

\alpha

\beta

) on multi-task learning using CTransPath and ABMIL.

$(w_{\text{NC}},w_{\text{HeC}},w_{\text{HoC}})$	(1,1,1)	(1,1.7,2)	(2,1.3,1)	(4,2,1)	(4,3,1)	(4,4,1)
Bal. Acc.	71.5	71.2	75.2	75.5	75.6	75.6
W. F1-Score	79.7	78.2	81.2	81.4	81.6	81.5

Table 5: Impact of WSD on classification loss weighting approach using CTransPath and ABMIL.

w_{\text{NC}}

w_{\text{HeC}}

and

w_{\text{HoC}}

are the weights given to slides with no consensus, heterogeneous consensus and homogeneous consensus.

Prostate Cancer Grading

See all 11 papers

Weakly Supervised Instance-Level Gleason Pattern Estimation Using Primary and Secondary Labels
Jul 26, 2026Nao Sugeta, Kaito Shiku, Shinnosuke Matsuo +1Weakly Supervised Learning Oncology
Kyushu University, Japan
GD-MIL: Grade-Disentangled Multiple Instance Learning for Multimodal Biochemical Recurrence Prediction in Prostate Cancer
Jun 8, 2026Dasari Naga RajuAdversarial Representation Learning Cancer Survival Prediction
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

Jul 26, 2026J/K move · Enter open · S save

Weakly Supervised Instance-Level Gleason Pattern Estimation Using Primary and Secondary Labels

Nao Sugeta, Kaito Shiku, Shinnosuke Matsuo, Ryoma Bise

Kyushu University, Japan

Weakly Supervised Learning Oncology Multiple Instance Learning Whole-Slide Image Analysis Automated Grading Prostate Cancer Grading

Open paper AskPDF arXiv

In prostate cancer histopathology, the Gleason Score is determined by the most frequent (Primary) and second most frequent (Secondary) Gleason patterns within a whole-slide image. Although these slide-level labels are routinely available in clinical practice, instance-level Gleason annotations are rarely provided, making patch-level learning challenging. We propose a Multiple Instance Learning (MIL) framework that estimates instance-level Gleason patterns from slide-level Primary and Secondary labels. The proposed method formulates instance-level learning according to the clinical definition of the Gleason Score by aggregating instance predictions into class counts and explicitly modeling the Primary pattern, Secondary pattern, and their dominance. Experimental results demonstrate that the proposed formulation enables effective instance-level learning and outperforms existing MIL approaches on the SICAP-MIL dataset.

Abstract

Prostate Cancer Grading

Explore similar work

Weakly Supervised Instance-Level Gleason Pattern Estimation Using Primary and Secondary Labels

GD-MIL: Grade-Disentangled Multiple Instance Learning for Multimodal Biochemical Recurrence Prediction in Prostate Cancer

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