Seize Disease includes many models stretching out into the field of computer vision so that histology is more easily accessible all around the world. Altering some of these models has shown to be beneficial in terms of accuracy and reliability. On top of that, the Seize Disease Website includes a full analyzing tool in which one can upload and alter an histology image. In addition to that, the use and introduction of the OpenFlexure microscope into the project is planned.
First, create a Conda or Pip environment for installing the necessary requirements. Once that's done, simply download the datasets and prepare them if necessary. You can find the links to the datasets in the references section.
# Using Pip
pip install -r requirements.txt
# Using Conda
conda create --name <env_name> --file requirements.txt
If you're using Google Colab, open any project file and make sure you have the requirements.txt file in it. Then, run the following command:
!pip install -r requirements.txt
first row → models shipped with this project
*changed/altered models
Custom Model X are new architectures, HistoSeg DP is also new, however, based on the HistoSeg Network
Multi-Class:
Dataset | Model | loss | dice score | pixel error | recall | precision | f1 | iou | jaccard index | year |
---|---|---|---|---|---|---|---|---|---|---|
Pannuke | HrNetV2 + OCR* | 1.2188 | 0.8814 | 0.0270 | 0.9047 | 0.9302 | 0.9172 | 0.7107 | 0.7892 | / |
... | SONNET | / | 0.824 | / | / | / | / | / | 0.686 | 2022 |
Binary:
Dataset | Model | loss | dice score | pixel error | recall | precision | f1 | iou | jaccard index |
---|---|---|---|---|---|---|---|---|---|
Pannuke | HrNetV2 + OCR* | 0.2217 | 0.8558 | 0.07797 | 0.9117 | 0.8098 | 0.8573 | 0.7510 | 0.7484 |
Edge:
Dataset | Model | loss | dice score | pixel error | recall | precision | f1 | iou | jaccard index |
---|---|---|---|---|---|---|---|---|---|
Pannuke | HrNetV2 + OCR* | 1.0987 | 0.9388 | 0.0141 | 0.9578 | 0.9578 | 0.9578 | 0.8612 | 0.8850 |
Dataset | Model | loss | accuracy | f1 | precision | recall | specificity at sensitivity |
---|---|---|---|---|---|---|---|
Pannuke | TinyVit | 0.0281 | 0.9955 | 0.9948 | 0.9966 | 0.9930 | 1.0 |
Tumor + Tissue
Dataset | Model | loss | accuracy | f1 | precision | recall | specificity at sensitivity | additional models | year |
---|---|---|---|---|---|---|---|---|---|
NCT-CRC-HE-100K | EfficientNetV2B2 (1/3) | 0.0096 | 0.9975 | 0.9975 | 0.9979 | 0.9972 | 1.0 | EfficientNetV2B1, EfficientNetV2B3 | 2021 |
... | Efficientnet-b0 | / | 0.9559 | 0.9748 | 0.9989 | / | 0.9945 | / | 2019 |
... | ResNeXt-50-32x4d | / | 0.9546 | 0.9746 | 0.9991 | / | 0.9943 | / | 2021 |
Dataset | Model | loss | accuracy | f1 | precision | recall | auc | specificity at sensitivity | additional models |
---|---|---|---|---|---|---|---|---|---|
Kather | ResAttInceptionV4* (1/2) | 0.9002 | 0.9271 | 0.9180 | 0.9449 | 0.8963 | 0.9970 | 0.9999 | ResAttInceptionV4* (smaller) |
Tumor
Dataset | Model | loss | accuracy | f1 | precision | recall | specificity at sensitivity | additonal models | Year |
---|---|---|---|---|---|---|---|---|---|
ICIAR2018_BACH_Challenge | TinyVit (1/3) | 0.5071 | 0.9000 | 0.9274 | 0.9583 | 0.8984 | 1.0 | EfficientNetV2B1, EffiencientNetV2B2 | 2022 |
... | Pretrained Resnet-101; Densenet-161 | / | 0.87 | / | / | / | / | / | 2018 |
Dataset | Model | loss | PSNR | SSIM | Model Config |
---|---|---|---|---|---|
Custom (Mixed Microscopy Images) | NafNet (256x256) | 0.0716 | 28.4709 | 0.837 | filters = 16, middle_block_num = 2, encoder_block_nums = (1,1,1,28), decoder_block_nums= = (1,1,1,1), block_type = NAFBLOCK, drop_out_rate = 0.0 |
... sligtly worse quality images | NafNet (128x128) | 0.0986 | 25.9457 | 0.7728 | filters = 32, middle_block_num = 1, encoder_block_nums = (1,1,1,7), decoder_block_nums= = (1,1,1,1), block_type = NAFBLOCK, drop_out_rate = 0.05 |
Dataset | Model | loss | PSNR | SSIM | Model Config |
---|---|---|---|---|---|
Custom (Histology Images) | NafNet (256x256) | 0.0647 | 30.9962 | 0.8412 | filters = 16, middle_block_num = 2, encoder_block_nums = (1,1,1,28), decoder_block_nums= = (1,1,1,1), block_type = NAFBLOCK, drop_out_rate = 0.0 |
... | NafNet (128x128) | 0.0587 | 29.622 | 0.8703 | filters = 32, middle_block_num = 1, encoder_block_nums = (1,1,1,7), decoder_block_nums= = (1,1,1,1), block_type = NAFBLOCK, drop_out_rate = 0.05 |
Dataset | Model | loss | PSNR | SSIM | Additions | Resolution | Additional Models |
---|---|---|---|---|---|---|---|
Custom (Mixed Microscopy Images) | HAT (small) | 0.0382 | 29.6236 | 0.8663 | Images were also slightly blurred, compressed and noised | 128 -> 256 | HAT (Mid) |
... | HAT (small) | 0.0272 | 26.7771 | 0.9039 | / | 64 -> 128 | / |
... | HAT (small) | 0.0623 | 23.7196 | 0.7859 | / | 64 -> 256 | / |
Binary Segmentation
Dataset | Model | loss | Dice/F1 | Recall | Precison | Additional Models | Model Config |
---|---|---|---|---|---|---|---|
Colorectal Adenocarcinoma Gland (CRAG) | HistoSeg* | 0.1567 | 0.8433 | 0.8018 | 0.8922 | VitaeV2 + OCR* | backbone = "xception" |
... | Custom Model L* | 0.2397 | 0.9033 | 0.8895 | 0.9176 | / | / |
... | Custom Model S/M* | 0.2724 | 0.854 | 0.8275 | 0.8849 | / | / |
... | HistoSeg DP* | 0.6872 | 0.7306 | 0.6843 | 0.7836 | / | based on HistoSeg mobilenetv2 |
---|---|---|---|---|---|---|---|
... | HistoSeg* | 0.5615 | 0.701 | 0.6589 | 0.7514 | / | backbone = "mobilenetv2" |
Binary Object Detection
Dataset | Model | loss |
---|---|---|
Colorectal Adenocarcinoma Gland (CRAG) | FasterRCNN RegNet Y 400MF | 0.8131 |
-
Kather Texture 2016 Image Tiles: Download here
-
ICIAR 2018 BACH: Download here
-
NCT-CRC-HE-100K: Download here
-
CPM 15, CPM 17: Download here
-
TNBC: Download here
-
Kumar: Download here
-
"Potato Tuber" included in Super Resolution/Restoration Dataset only : Download here
-
Malaria Bounding Boxes included in Super Resolution/Restoration Dataset only: Download here
-
Efficient NetV2: https://github.com/leondgarse/keras_efficientnet_v2 @leondgarse
-
Data Preparation for Merged Dataset: https://github.com/nauyan/NucleiSegmentation @ nauyan
-
Data Preparation for PanNuke: https://github.com/Mr-TalhaIlyas/Prerpcessing-PanNuke-Nuclei-Instance-Segmentation-Dataset @Mr-TalhaIlyas
-
Mish Activation Function: https://www.kaggle.com/code/imokuri/mish-activation-function/notebook @SUGIYAMA Yoshio
-
Segment Anything: https://segment-anything.com/ @Meta AI, Github
-
StarDist: https://github.com/stardist/stardist @StarDist
-
NafNet/Restorer and SR Losses: https://github.com/TrystAI/restorers @TrystAI
-
HistoSeg: https://github.com/saadwazir/HistoSeg @saadwazir
-
Efficient Attention: https://github.com/cmsflash/efficient-attention @Shen Zhuoran
-
Pannuke Dataset Comparisions: https://paperswithcode.com/dataset/pannuke | SONNET
-
NCT-CRC-HE-100K Dataset Comparisions: https://paperswithcode.com/sota/medical-image-classification-on-nct-crc-he | EfficientNet
-
ICIAR 2018 BACH Dataset Comparisions: https://figshare.com/articles/dataset/Results_on_the_ICIAR_2018_challenge_dataset_/19676718/1; https://arxiv.org/pdf/1808.04277.pdf