Benchmarking Self-Supervised Learning on Diverse Pathology Datasets

We execute the largest-scale study of SSL pre-training on pathology image data. Our study is conducted using 4 representative SSL methods below on diverse downstream tasks. We establish that large-scale domain-aligned pre-training in pathology consistently out-performs ImageNet pre-training.

Pre-trained weights

1. bt_rn50_ep200.torch: ResNet50 pre-trained using Barlow Twins
2. mocov2_rn50_ep200.torch: ResNet50 pre-trained using MoCoV2
3. swav_rn50_ep200.torch: ResNet50 pre-trained using SwAV
4. dino_small_patch_${patch_size}_ep200.torch: ViT-Small/${patch_size} pre-trained using DINO

md5sum

Image statistics

We used the following statistics for image intensity standardization (normalization):

mean: [ 0.70322989, 0.53606487, 0.66096631 ]
std: [ 0.21716536, 0.26081574, 0.20723464 ]

which are values corresponding to R, G, and B channels respectively, determined from 10% of the training samples.