💻 Medical Image AI Challenge 2023 : Pathology data

Intro

• This competition is a medical artificial intelligence competition hosted by Seoul National University Hospital
• its purpose is to predict disease risk (survival/recurrence) using pathological images and clinical information of malignant melanoma patients.
• 12 out of 49 teams got advanced to the finals, and unfortunately our team did not make it into the rankings in private data, despite we ranked 4th in public data.

Team Members

박성현	박혜나	신건희	신중현	한상준

main concept

• Whole Slide Image (WSI)
  • The entire shape of the tissue contained on a glass slide is converted into a high-resolution digital image in a short period of time.
  • In general, the size of the image is 100,000 x 100,000 pixels or more, so in this competition we use images reduced and saved at 100x magnification for smooth processiong.
• Multiple Instance Learning (MIL)
  • One of the weakly supervised learning methods, the deep learning model automatically selects k patches that are deemed useful for label prediction.
  • If even one lesion exists in WSI, it is classified as a positive slide, and if none of the lesions exist, it is classified as a negative slide

Process

1. EDA (image data / table data)
2. baseline code construction & modulation
3. search references (models and WSI techniques)
4. patch extraction experiment
5. dataloader experiment (save patch to pickle)
6. train model (AB_MIL / DSMIL / SimCLR + tabNet / AC_MIL / MHIM_MIL)
7. Ensemble

Dataset

• number of images : 894
• number of patientID : 217
• Image
  • H&E stained pathology whole slide images(WSI)
  • width : (avg) 28,440 / (max) 54,945 / (min) 8,963
  • height : (avg) 19,106 / (max) 24,538 / (min) 5,656
• Table
  • data of each WSI images composed by 18 columns
  • columns : 'Slide_name', 'Patient_ID', 'Recurrence', 'Location', 'Diagnosis', 'Growth phase', 'Size of tumor', 'Depth of invasion', 'Level of invasion', 'Mitosis', 'Histologic subtype', 'Tumor cell type', 'Surgical margin', 'Lymph node', 'Breslow thickness', 'Precursor lesion', 'Date_of_diagnosis', 'Date_of_recurrence'
  • Recurrence : 0 or 1 (nonrecurrent or recurrent)
    • (WSI proportion) recurrent : 688 | nonrrent : 206
    • (patientID proportion)recurrent : 48 | nonrrent : 169

File

    ├─── EDA
    │   ├── EDA.ipynb
    │─── data_process
    │   ├── split_dataset.ipynb
    │   ├── dataset2pkl_train.py
    │   ├── dataset2pkl_test_public.py
    │─── check_status
    │   ├── check_capacity.ipynb
    │   ├── check_file_status.ipynb
    │   ├── check_fold_patches.ipynb
    │   ├── check_gpu.ipynb
    ├─── AB_MIL
    │─── DSMIL
    ├─── TabNet
    │   ├── app
    │   │   ├── lib
    │   │   ├── modules
    │─── MHIM_MIL
    │   ├── camelyon16
    │   ├── modules

평가 Metric

• AUROC for recurrence prediction on a per-patient basis

Reference papers

• A Simple Framework for Contrastive Learning of Visual Representations
• Attention-based Deep Multiple Instance Learning
• Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification
• Interpretable Prediction of Lung Squamous Cell Carcinoma Recurrence With Self-supervised Learning
• Dual-stream Multiple Instance Learning Network for Whole Slide Image Classification with Self-supervised Contrastive Learning
• Multiple Instance Learning Framework with Masked Hard Instance Mining for Whole Slide Image Classification
• TransMIL_Transformer based Correlated Multiple Instance Learning for Whole Slide Image Classification