Recommended MONAI MIL baseline for variable patient-level histopathology bags

[motazkhasawneh]

Hello MONAI community. I am considering MONAI for an efficient patient-level histopathology classification baseline and would appreciate guidance on the most appropriate supported components.

The general setup is public H&E images, multiple slides per patient, variable numbers of crops, and patient-level labels. The goal is one prediction per patient under a limited inference budget. I am interested in comparing a compact trainable encoder plus MIL with frozen pathology-pretrained embeddings plus a small MIL head, while keeping patient and tissue-source groups separate across evaluation folds.

I found the PANDA MIL example: https://github.com/Project-MONAI/tutorials/blob/main/pathology/multiple_instance_learning/panda_mil_train_evaluate_pytorch_gpu.py

Could you point me to:

1. A recommended MONAI example or pattern for extending slide-level MIL to patient-level bags across multiple slides?
2. Supported ways to handle variable bag sizes, padding/masking, and chunked feature extraction without allowing padded crops to affect aggregation?
3. Suitable MONAI MIL aggregators for a fair comparison between frozen embeddings and compact end-to-end training, including any important limitations?

This is a design question, not a bug report. Links to examples, documentation, or relevant papers would be very helpful. I can discuss the general setup, but unpublished methods, results, and research artifacts will remain private. Thank you.

[alleyxd]

Some pointers from reading monai/networks/nets/milmodel.py, which the PANDA example uses:

1. Patient-level bags. MILModel doesn't know about slides. It takes [B, N, C, H, W] and treats the N tiles as one bag. So patient level is just a data question: pool the tiles from all of a patient's slides into one bag (the PANDA dataloader samples a fixed tile_count per slide, and you'd sample per patient instead).

2. Variable bag sizes. There's no mask argument. forward reshapes to B*N tiles and back, so every item in a batch needs the same N, and padded tiles would enter the mean/attention. The practical options are:

   • sample a fixed N per bag (with replacement for small bags), which is what PANDA does
   • use batch size 1 for exact variable bags
   • call model(x, no_head=True) to get the [B, N, D] tile features and write a small masked attention head yourself (set masked logits to -inf before the softmax)

3. Frozen embeddings vs end-to-end. You can reuse the same aggregators on precomputed features by passing them as [B, N, D, 1, 1] with backbone=torch.nn.Flatten() and backbone_num_features=D. The reshape then just passes them through. Custom backbones support mean, max, att and att_trans, but not att_trans_pyramid, which needs the ResNet layer hooks. Using the same mil_mode and the same bag sampling on both sides keeps the comparison fair.

For the attention pooling itself, the reference is Ilse et al., "Attention-based Deep Multiple Instance Learning" (2018). att_trans adds a transformer over the tiles before pooling (Myronenko et al., MICCAI 2021).