A University College London MSc Health Data Science Dissertation and AstraZeneca Summer Studentship Project
Project: Human-like AI Analysis on Pathologic Images
Dataset Used: https://www.bracs.icar.cnr.it/
- To explore the effects of human-like image transformation including magnification and colour space on the performance of AI and ML analysis of pathologic whole slide images (WSI)
- To design a computationally efficient ML model for WSI analysis combining various ML techniques
- To create reusable assets for pathologic image analysis and discuss implications on study results
DATA_PATH - BRACS_WSI
├── train
| ├── Group_AT
| | ├── Type_ADH
| | └── Type_FEA
| ├── Group_BT
| | ├── Type_N
| | ├── Type_PB
| | └── Type_UDH
| └── Group_MT
| ├── Type_DCIS
| └── Type_IC
├── val
. .
└── test
.
ANNOT_PATH - BRACS_WSI
├── train
| ├── Group_AT
| | ├── Type_ADH ─── BRACS_<ID1>.qpdata, BRACS_<ID2>.qpdata, ...
| | └── Type_FEA ...
| ├── Group_BT
| | ├── Type_N ...
| | ├── Type_PB ...
| | └── Type_UDH ...
| └── Group_MT
| ├── Type_DCIS ...
| └── Type_IC ...
├── val
. .
└── test
.
PROJECT_PATH
├── data
| ├── norms
| | └── <normalisation_values>.npy ...
| ├── roi_patches
| | ├── sample ...
| | ├── train ...
| | └── val ...
| ├── roi_test_imgs (ROI only test set images from BRACS)
| | ├── 0_N ...
| | ├── 1_PB ...
| | .
| | └── 6_IC ...
| └── slide_patches
| ├── sample ...
| ├── train ...
| └── val ...
├── models
| ├── train_obj_detector ...
| └── train_patch_classifier ...
|
└── output
└── plots
├── evaluation
.
├── train_obj_detector
└── train_patch_classifier
The directories structure from BRACS should be identical to above by default. Otherwise, format and align with above.
- open QuPath to edit the
qp_annotations_to_json.groovyscript - specify the
ANNOT_PATHin line 5def outputPath = "path/to/BRACS_WSI_Annotations" - run and generate
.geojsonfiles from the original.qpdatafiles into the sameANNOT_PATH
Most Python scripts includes required arguments, use python <script>.py -h to read the details
python src/extract_patches.py -p <project_path> -r /some/path/BRACS/BRACS_WSI/ -a /some/path/BRACS_WSI_Annotations/
python src/extract_slides.py -p <project_path> -r /some/path/BRACS/BRACS_WSI/ -a /some/path/BRACS_WSI_Annotations/python src/patch_transform_showcase.py -p <project_path>python src/train_obj_detector.py --project_root <project_path> --color_space <color_space>
python src/test_obj_detector.py -p <project_path> -r /some/path/BRACS/BRACS_WSI/ -a /some/path/BRACS_WSI_Annotations/ -c <color_space> -k <top_k_boxes> -n <nms_threshold>python src/train_patch_classifier.py --project_root <project_path> --color_space <color_space> --mag <magnification>
python src/test_patch_classifier.py -p <project_path> -c <color_space> -m <magnification>python src/run_inference.py -p <project_path> -r /some/path/BRACS/BRACS_WSI/ -a /some/path/BRACS_WSI_Annotations/ -s <slide_tile_size> -t <roi_tile_size> -c <classifier_colour_space> -o <obj_detect_colour_space> -m <classifier_magnification> -b <classifier_batch_size> -k <top_k_boxes> -n <nms_threshold>