Closed dmhdmhdmh closed 3 weeks ago
In [1][2][3], for several methods, they achieve similar performance by utilizing the features obtained from ResNet-50 and the features obtained from PLIP. [1] Feature Re-Embedding: Towards Foundation Model-Level Performance in Computational Pathology [2] Towards A Generalizable Pathology Foundation Model via Unified Knowledge Distillation [3] A Multimodal Knowledge-enhanced Whole-slide Pathology Foundation Model
For the ResNet-50, Max-Pooling only outperforms TransMIL in BRACS-7* and BRACS-7, and achieves the lowest performance in the Mean results. In the NSCLS-2, Max-Pooling outperforms most of the comparative methods, and Mamba2MIL (recent work: https://www.arxiv.org/pdf/2408.15032) gets the similar conclusions, where the Max-Pooling gets the third highest performance in the NSCLS dataset.
I hope that you can directly repeat my experiments by using provided codes before criticizing the rigor and accuracy of my experiments in the paper.
In the Cancer Subtyping experiment, the training results using features extracted by PLIP seem even worse than those using features extracted by Resnet50. Moreover, I observed that simply using Max-Pooling appears to outperform most of the comparative methods. I want to ask if you truly conducted your experiments with the necessary rigor and seriousness?