Abstract:
Computer aided diagnosis is widely used in medical imaging for the diagnosis of many
diseases such as cardiomegaly, brain and kidney tumor, lung cancer, COVID-19 and
may more. For the past few decades, computer aided diagnosis has significantly
improved due to the development of better architecture used for the diagnosis. Knee
injury diagnosis using deep learning techniques is highly popular due its high detection
rate and is highly localized. Many state-of-the-art-deep learning models have been
used for the detection of abnormalities, meniscus tear and ACL tears in Knee MRI
scans. These models include RESNET, Google-Net, VGG19 and VGG16, Alex-Net
and many other, all giving significant results. In this study we used a custom 3D CNN
model which is light in weight. For training we are using two datasets, one provided
by Stanford ML group and the other form Hospital in Croatia. We combined the two
dataset and split it into 80-20 ration (80% of the data used for training and remaining
for testing purposes). Both the dataset has extreme class imbalance, so we used data
augmentation and class weights to rectify its effect on the training process. Further the
voxel intensities for the two datasets were different (one dataset was in 8-bit format
and the second was in 12-bit format), so we normalized the intensity values using
mathematical formulas. For contrast, we performed adaptive histogram equalization
Average accuracy and AUC achieved by our model on training set is 97.6 and 99.3
respectively, during 5-fold cross validation.
