README.md

metadata

language: en
license: mit
datasets:
  - DDR
  - FGADR
  - IDRID
  - MESSIDOR
  - RETLES
library: torchSeg
model-index:
  - name: unet_seresnext50_32x4d
    results:
      - task:
          type: image-segmentation
        dataset:
          name: IDRID
          type: IDRID
        metrics:
          - type: roc_auc
            value: 0.6701094508171082
            name: AUC Precision Recall - IDRID COTTON_WOOL_SPOT - COTTON_WOOL_SPOT
          - type: roc_auc
            value: 0.7860875129699707
            name: AUC Precision Recall - IDRID EXUDATES - EXUDATES
          - type: roc_auc
            value: 0.6743975877761841
            name: AUC Precision Recall - IDRID HEMORRHAGES - HEMORRHAGES
          - type: roc_auc
            value: 0.39846163988113403
            name: AUC Precision Recall - IDRID MICROANEURYSMS - MICROANEURYSMS
      - task:
          type: image-segmentation
        dataset:
          name: FGADR
          type: FGADR
        metrics:
          - type: roc_auc
            value: 0.4449217915534973
            name: AUC Precision Recall - FGADR COTTON_WOOL_SPOT - COTTON_WOOL_SPOT
          - type: roc_auc
            value: 0.6951484084129333
            name: AUC Precision Recall - FGADR EXUDATES - EXUDATES
          - type: roc_auc
            value: 0.6508341431617737
            name: AUC Precision Recall - FGADR HEMORRHAGES - HEMORRHAGES
          - type: roc_auc
            value: 0.2895563244819641
            name: AUC Precision Recall - FGADR MICROANEURYSMS - MICROANEURYSMS
      - task:
          type: image-segmentation
        dataset:
          name: MESSIDOR
          type: MESSIDOR
        metrics:
          - type: roc_auc
            value: 0.3307325839996338
            name: >-
              AUC Precision Recall - MESSIDOR COTTON_WOOL_SPOT -
              COTTON_WOOL_SPOT
          - type: roc_auc
            value: 0.7123324871063232
            name: AUC Precision Recall - MESSIDOR EXUDATES - EXUDATES
          - type: roc_auc
            value: 0.3926454186439514
            name: AUC Precision Recall - MESSIDOR HEMORRHAGES - HEMORRHAGES
          - type: roc_auc
            value: 0.4098129868507385
            name: AUC Precision Recall - MESSIDOR MICROANEURYSMS - MICROANEURYSMS
      - task:
          type: image-segmentation
        dataset:
          name: DDR
          type: DDR
        metrics:
          - type: roc_auc
            value: 0.5084977746009827
            name: AUC Precision Recall - DDR COTTON_WOOL_SPOT - COTTON_WOOL_SPOT
          - type: roc_auc
            value: 0.6117375493049622
            name: AUC Precision Recall - DDR EXUDATES - EXUDATES
          - type: roc_auc
            value: 0.5447860956192017
            name: AUC Precision Recall - DDR HEMORRHAGES - HEMORRHAGES
          - type: roc_auc
            value: 0.23405438661575317
            name: AUC Precision Recall - DDR MICROANEURYSMS - MICROANEURYSMS
      - task:
          type: image-segmentation
        dataset:
          name: RETLES
          type: RETLES
        metrics:
          - type: roc_auc
            value: 0.5254419445991516
            name: AUC Precision Recall - RETLES COTTON_WOOL_SPOT - COTTON_WOOL_SPOT
          - type: roc_auc
            value: 0.7039055824279785
            name: AUC Precision Recall - RETLES EXUDATES - EXUDATES
          - type: roc_auc
            value: 0.5196094512939453
            name: AUC Precision Recall - RETLES HEMORRHAGES - HEMORRHAGES
          - type: roc_auc
            value: 0.4127877354621887
            name: AUC Precision Recall - RETLES MICROANEURYSMS - MICROANEURYSMS

Lesions Segmentation in Fundus

Introduction

We focus on the semantic segmentations of:

1. Cotton Wool Spot
2. Exudates
3. Hemmorrhages
4. Microaneurysms

For an easier use of the models, we refer to cleaned-up version of the code provided in the fundus lesions toolkit.

Architecture

The model uses unet_seresnext50_32x4d as architecture. The implementation is taken from torchSeg

Training datasets

The model was trained on the following datasets: DDR, FGADR, IDRID, MESSIDOR, RETLES

Resolution

The image resolution for training was set to 1024 x 1024.