Medical Image Analysis generally refers to the technique of creating visual representations of the interior of a body. This helps to understand how tissues or organs function. It reveals internal body structures and helps medical practitioners properly diagnose diseases. Deep Learning (DL) systems like Convolutional Neural Networks (CNN) can help in presenting a hierarchical representation of these images.
Medical Image Segmentation
The method of identifying organs or lesions from CT scans or MRI images and delivering essential information about the shapes and volumes of these organs is called Medical Image Segmentation. CNNs use a 2D input image and apply 2D filters. It requires minor preprocessed data and discovers representations in a self-taught manner. Deep neural networks can find out hierarchical feature representations in such a way that the higher level features can be derived from the lower level features. Thus, improvements in computer vision has helped in image segmentation, image fusion, image annotation, diagnosis & prognosis, lesion detection, etc.
Deep Learning in detection of Alzheimer’s and Parkinson’s disease
CNN has the ability to handle imbalanced datasets and provide unbiased results. A CNN architecture consists of: a convolution layer, a pooling layer and fully connected layers.
Most of the ischemic brain stroke computer-aided diagnosis (CAD) systems are developed using clustering and Support Vector Machines (SVM) and random forest classifiers. Specialized neurologists may not be available in remote villages. For those cases, signals/images of patients are transmitted to a cloud where a ML-trained CAD can provide the diagnosis. After that, the outcome of the model can be sent to the mobile phone of a non-expert clinician for a preliminary diagnosis.
Detection of Tumor, Skin Cancer & Diabetic Retinopathy
Deep Learning models are trained using supervised learning. The input data of which comprises of dermoscopic images of skin diseases and it’s corresponding target output labels are: skin disease classes like ‘benign’ or ‘malignant’. A DL algorithm can predict tumor proliferation by using a tumor probability heatmap. It classifies the tumor probability of overlapping tissue patches. Especially, the final images produced clearly show the features of the tumor including it’s shape, location, density. With the help of deep learning, we can enable the automation of progress monitoring. Earlier manual detection of diabetic retinopathy was a time-consuming process. Now automated detection based on deep learning models has proven their better accuracy.
Thus, we see that deep Learning holds the potential to completely transform the healthcare domain. It has undoubtedly become an integral part of the medical industry today. These computational modeling techniques for image analysis has great impact on scientific research as well as clinical applications. Due to it’s black-box like characteristics of the models, comprehending the intuition of the learned models is still a challenge.
