Micro Suturing Effectualness

Problem Statement:

Design a vision-based techniques to evaluate effectualness of micro suturing by trainee neurosurgeons.

Parameters:

In total 6 parameters were decided, to evaluate a given suture. This github repo focuses on 2 parameters to evaluate "knots" in suture.
Namely,

Alignment of knots
Uniformity in distance between knots

Objectives:

Detection of knots in suturing images.
Perform parametric evaluation for effectualness for micro-suturing.

Preprocessing:

One of the major steps was preprocessing, as the suture images had noise and illumination effect which needed to dealt with. Mode Clipping of images was done which reduced noise, illumination effect and good segregation of background and foreground.
In order to reduce the search space for knots, segmentation and gaussian model based noise reduction followed by contour hunting was done. Which reduced to search space drastically.

Knot Detection:

We wanted a intensity invariant classifier, to achieve this a SVM classifier was trained on HOG feature descriptors of knots and non-knots patches. In order to improve the performance, generative algorithms like GDA and GMM were also implemented as a post processing step.

Code:

To train a SVM classifier on HOG feature descriptors of knots and non knots patches.

python svm_extra_features.py

To train GDA classifier.

python biclass_gda_with_save.py

To train GMM classifier.

python gmm_classifier.py

To generate prediction boxes on test set (SVM, NMS, GDA)

python svm+nms+gda_boxes_save.py

To generate prediction boxes on test set (SVM, NMS, GMM)

python generate_prediction_boxes.py

To test on given single test image

python main.py

Evaluated test images

python evaluation.py