Collaborative Learning with Curriculum Loss for Accurate and InterpretableWeakly Supervised Whole-Slide Image Classification Collaborative Learning with Curriculum Loss for Whole-Slide Image Classification

Yang Rui, Liu Pei, Ji Luping
Proceedings of 2024 International Conference on Machine Learning and Intelligent Computing, PMLR 245:33-39, 2024.

Abstract

Weakly-Supervised Learning (WSL) has been increasingly concerned in Whole-Slide Image (WSI) classification, meanwhile, an open question arises: could WSL-based models provide us with an accurate interpretation of their decisions? Although many research works have made exciting progress via building an Auxiliary Instance Branch (AIB) on a bag-level network, there are still two typical problems to be confronted with in training WSL-based AIB: i) an overwhelming influence of negative instances and ii) the inconsistent learning between bag-level network and AIB. To address them, this paper proposes collaborative learning with curriculum loss. This scheme, on one hand, provides a curriculum loss for optimizing AIB, to alleviate the first problem. Considering the knowledge reliability in WSL, this loss generalizes an original quality focal loss to WSL scenarios by curriculum instances. On the other hand, to overcome the second problem, this scheme trains a bag-level network under the supervision of AIB by a reversed curriculum loss, making both learn collaboratively. Comparative experiments prove that our scheme could often surpass existing ones in both accuracy and interpretability. Moreover, it is found that the knowledge reliability-inspired curriculum instance is a critical factor in bringing comprehensive improvements.

Cite this Paper

BibTeX
@InProceedings{pmlr-v245-rui24a, title = {Collaborative Learning with Curriculum Loss for Accurate and InterpretableWeakly Supervised Whole-Slide Image Classification Collaborative Learning with Curriculum Loss for Whole-Slide Image Classification}, author = {Rui, Yang and Pei, Liu and Luping, Ji}, booktitle = {Proceedings of 2024 International Conference on Machine Learning and Intelligent Computing}, pages = {33--39}, year = {2024}, editor = {Nianyin, Zeng and Pachori, Ram Bilas}, volume = {245}, series = {Proceedings of Machine Learning Research}, month = {26--28 Apr}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v245/main/assets/rui24a/rui24a.pdf}, url = {https://proceedings.mlr.press/v245/rui24a.html}, abstract = {Weakly-Supervised Learning (WSL) has been increasingly concerned in Whole-Slide Image (WSI) classification, meanwhile, an open question arises: could WSL-based models provide us with an accurate interpretation of their decisions? Although many research works have made exciting progress via building an Auxiliary Instance Branch (AIB) on a bag-level network, there are still two typical problems to be confronted with in training WSL-based AIB: i) an overwhelming influence of negative instances and ii) the inconsistent learning between bag-level network and AIB. To address them, this paper proposes collaborative learning with curriculum loss. This scheme, on one hand, provides a curriculum loss for optimizing AIB, to alleviate the first problem. Considering the knowledge reliability in WSL, this loss generalizes an original quality focal loss to WSL scenarios by curriculum instances. On the other hand, to overcome the second problem, this scheme trains a bag-level network under the supervision of AIB by a reversed curriculum loss, making both learn collaboratively. Comparative experiments prove that our scheme could often surpass existing ones in both accuracy and interpretability. Moreover, it is found that the knowledge reliability-inspired curriculum instance is a critical factor in bringing comprehensive improvements.} }
Endnote
%0 Conference Paper %T Collaborative Learning with Curriculum Loss for Accurate and InterpretableWeakly Supervised Whole-Slide Image Classification Collaborative Learning with Curriculum Loss for Whole-Slide Image Classification %A Yang Rui %A Liu Pei %A Ji Luping %B Proceedings of 2024 International Conference on Machine Learning and Intelligent Computing %C Proceedings of Machine Learning Research %D 2024 %E Zeng Nianyin %E Ram Bilas Pachori %F pmlr-v245-rui24a %I PMLR %P 33--39 %U https://proceedings.mlr.press/v245/rui24a.html %V 245 %X Weakly-Supervised Learning (WSL) has been increasingly concerned in Whole-Slide Image (WSI) classification, meanwhile, an open question arises: could WSL-based models provide us with an accurate interpretation of their decisions? Although many research works have made exciting progress via building an Auxiliary Instance Branch (AIB) on a bag-level network, there are still two typical problems to be confronted with in training WSL-based AIB: i) an overwhelming influence of negative instances and ii) the inconsistent learning between bag-level network and AIB. To address them, this paper proposes collaborative learning with curriculum loss. This scheme, on one hand, provides a curriculum loss for optimizing AIB, to alleviate the first problem. Considering the knowledge reliability in WSL, this loss generalizes an original quality focal loss to WSL scenarios by curriculum instances. On the other hand, to overcome the second problem, this scheme trains a bag-level network under the supervision of AIB by a reversed curriculum loss, making both learn collaboratively. Comparative experiments prove that our scheme could often surpass existing ones in both accuracy and interpretability. Moreover, it is found that the knowledge reliability-inspired curriculum instance is a critical factor in bringing comprehensive improvements.
APA
Rui, Y., Pei, L. & Luping, J.. (2024). Collaborative Learning with Curriculum Loss for Accurate and InterpretableWeakly Supervised Whole-Slide Image Classification Collaborative Learning with Curriculum Loss for Whole-Slide Image Classification. Proceedings of 2024 International Conference on Machine Learning and Intelligent Computing, in Proceedings of Machine Learning Research 245:33-39 Available from https://proceedings.mlr.press/v245/rui24a.html.

Related Material