melSeg

This is a tool to train semantic-segmentation model targeting skin leison segmentation in digital images.

Install requirements

pip install -r requirements.txt

Download and structure HAM10000 Dataset

• Download Link: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/DBW86T
• Unzip all zipped directories from the download.
• Create a directory called HAM10000 with two sub-directories: images, masks
• Move all RGB-images to the sub-directory: HAM10000/images
• Move all Mask-images to the sub-directory: : HAM10000/masks

Run Training

This pipeline starts training a semantic-segmentation model.
Training-parameters can be seen or, modified from training_config.yaml.

Check help

python3 melseg_trainer.py --help

Command example

python3 melseg_trainer.py --config training_config.yaml

Run Analysis

This pipeline analyzes the training-logs [.csv] of a cross-validation run, by:

• finding the best performance of each fold using minimum loss and maximum epoch
• and, calculating average of fold-wise best performances. This pipeline saves the analysis report as a [.csv] file.

Check help

python3 melseg_analyzer.py --help

Command example

python3 melseg_analyzer.py --root example_logs/ --outpath example_logs/log_analysis.csv

Split Config

This pipeline generates individual sweeps from a training-config.

Check help

python3 config_splitter.py --help

Command example

python3 config_splitter.py --config training_config.yaml --outdir example_configs

Code Formatting

Black is a Python code formatter that enforces a consistent style by rewriting code to conform to PEP 8 style guide.

Command

sh run_black.sh

Code Linting

Linting is the process of static code analysis that examines code for errors, stylistic issues, and potential problems. There are many different linters available for Python, Flake8 is used here.

Command

sh run_flake8.sh

Metric Calculation

Following metrics are calculated while training the model.

PRECISION

Precision score is the number of true positive results divided by the number of all positive results.

$$Precision=TP/(TP+FP)$$

RECALL

Recall score, also known as Sensitivity or true positive rate, is the number of true positive results divided by the number of all samples that should have been identified as positive.

$$Recall=TP/(TP+FN)$$

ACCURACY

Accuracy score, also known as Rand index is the number of correct predictions, consisting of correct positive and negative predictions divided by the total number of predictions.

$$Accuracy=(TP+TN)/(TP+TN+FP+FN)$$

DICE COEFFICIENT (F1-SCORE)

Dice coefficient, also known as F1 score is the harmonic mean of precision and recall. In other words, it is calculated by 2 x intersection divided by the total number of pixel in both images. $$Dice=2TP/(2TP+FP+FN)$$

INTERSECTION OVER UNION (IoU)

Intersection over Union (IoU), also known as Jaccard index is the area of the intersection over union of the predicted segmentation and the ground truth. $$IoU=TP/(TP+FP+FN)$$

Ref: https://www.kaggle.com/code/nghihuynh/understanding-evaluation-metrics-in-segmentation/notebook