4. DL-based systems in clinical diagnosis
Several studies observed that DL improved accuracy in diagnosis and clinical assessment. The DL architectures performed well, especially in cardiovascular MRI (CMR), fatty liver disease (FLD) tissue characterization and rheumatic arthritis. Although these are separate areas of medical science where different modalities are used, the success of DL motivates the expansion of its horizon. Moreover, the success of DL-based strategy in anatomical research inspire application in CVD diagnosis and prediction. Schmauch et al. proposed a DenseNet-based supervised learning system for the detection classification of focal liver lesions (FLL) [124]. The authors demonstrated region-of-convergence (ROC) and area under the curve (AUC) scores of 0.935 and 0.916 for lesion detection and classification on 367 ultrasound images [124]. Huynh et al. [125] applied transfer learning for detecting malignant tumours. Features from pre-trained AlexNet were used for training. The network had performance similar to analytical methods (AUC = 0.81) [125]. Hatipoglu et al. performed cell segmentation using three DL-based methods: CNN, DBN and autoencoder. The comparative study reported that CNN and autoencoders outperformed conventional segmentation techniques [126].
Table III Benchmarking deep learning-based studies in clinical diagnosis.