Python-based project performing binary segmentation on PH2 dataset : https://www.dropbox.com/s/k88qukc20ljnbuo/PH2Dataset.rar?file_subpath=%2FPH2Dataset
This project generated binary masks from dermatological samples fed into the system which are ultimately classified into three categorical divisions.
The motivation to achieve enhanced accuracy on predictions made by different Deep Learning techniques - Segmentation and Classification. Sometimes, CNNs when fed with three channel images fail to learn structural information necessary to classify them. Annotated training data with boundaries clearly defined enables efficient detection and learning. Segmentation network used in medical imaging generates binary masks which when fed to a classification network performs much better. For this project, popular PH2 dataset was employed to verify segmentation, followed by classificaton between common_nevus, atypical_nevus and melanoma types of dermatology samples of affected tissue. For feature extraction, Efficient-net-b4-widese is used with their pre-final layer modified for the need of the objective. As the network is trained and evaluated on annotated masks, model weights are saved having the best performance w.r.t. the loss function employed. Consequently, the test dataset is fed to the system to generate masks followed by classifier network classfiying them.
The increasing incidence of melanoma has recently promoted the development of computer-aided diagnosis systems for the classification of dermoscopic images. The PH² dataset has been developed for research and benchmarking purposes, in order to facilitate comparative studies on both segmentation and classification algorithms of dermoscopic images. PH² is a dermoscopic image database acquired at the Dermatology Service of Hospital Pedro Hispano, Matosinhos, Portugal.
The dataset is available at:
https://web.inf.ufpr.br/vri/databases/breast-cancer-histopathological-database-breakhis/
In this project, the histopathological image samples of human breast tissue have been classified into two categories
Common NevusAtypical NevusMelanoma
Network Architectures Used:
- Segmentation:
U-Net - Classifier:
EfficientNet-b4-widese
Three types of classifiers are employed for fitness evaluation
Support Vector Machines (kernel<--rbf)K Nearest Neighbours (neighbours<--2)Multi-layer Perceptron
-
Roadmap
Since the entire project is based on Python programming language, it is necessary to have Python installed in the system. It is recommended to use Python with version >=3.9.
The Python packages which are in use in this project are matplotlib, numpy, pandas,scikit-learn, torch and torchvision. All these dependencies can be installed just by the following command line argument
pip install requirements.txt
Current directory ----> data
|
|
|
------------------> train
| |
| -------------------------
| | |
| V V
| input target
| | |
| ------------------ ------------------
| | | | |
| V V V V
| image_1 .... image_n mask_1 .... mask_n
|
|
|
|
------------------> val
| |
| -------------------------
| | |
| V V
| input target
| | |
| ------------------ ------------------
| | | | |
| V V V V
| image_1 .... image_n mask_1 .... mask_n
|
|
|
|
|
------------------> train_classifier
| |
| -------------------------
| | | |
| V V V
| gen_mask 1 gen_mask 2 ... gen_mask n
|
|
|
------------------> val_classifier
| |
| -------------------------
| | | |
| V V V
| gen_mask 1 gen_mask 2 ... gen_mask n
|
|
------------------> test
|
-------------------
| |
V V
image_1 .... image_n
SVM: 'SVM'
MLP: 'MLP'
KNN: 'KNN'