Aims:
Diagnosis of scoliosis requires periodic detection, and as frequent exposure to radiative imaging may cause cancer. A safer and economic alternative imaging modality is being explored such as ultrasound imaging (US) Unlike other radioactive modalities, an US image is full of speckle and occlusion noise which often supress the useful information of that image. Through our research, we propose a novel AI architecture for a bony feature detection which can be further used for scoliosis assessment.
Method:
Among popular AI techniques, Convolutional Neural Networks (CNNs) is well proven for analyzing huge number of medical image data. But due to the poor quality of US images, the CNN architecture often misses or misrepresents regions of interest in the image. Through our research, we are proposing a novel CNN architecture incorporating (a) improvised Inception block (b) newly designed decoder side skip pathways into the basic U-Net architecture. This will help overcome the current shortfalls of CNN in handling Ultrasound images.
Result:
The proposed model is tested on 109 spine ultrasound image data sets received from Hong Kong Polytechnic University. Our method is evaluated using the popular Jaccard Index and Dice Coefficient which measures the similarity with the original ground truth and is compared with (a) the basic U-net segmentation model and (b) one state-of-the-art model i.e. MultiResUNet. The research shows that our proposed model gives a clearer segmentation output, vis-a-vie both the models, especially in the important Regions of Interest such as upper and lower rib. Also, our segmentation output gives a better average Jaccard score and Dice score.
Conclusion:
Our research shows an early indication that this hybridized CNN is a promising approach and can meet the objectives for a scoliosis assessment using US image.