Call For Papers (CFP) 

Deep learning (DL) methods have led to significant advances in various applications. In the medical domain, DL techniques have been used for processing and analyzing a broad spectrum of medical image modalities as they provide valuable information on understanding, detecting, and diagnosing diseases. Some examples of medical image modalities include chest X-rays, computed tomography, magnetic resonance imaging, fundus images, colposcopic images, ultrasound echocardiography, and microscopic images. DL-based image analysis can help doctors and researchers in several aspects, such as segmenting organs-of-interest or clinically important regions, localizing and measuring disease manifestations, identifying and classifying disease patterns, enhancing image quality for better/easier manual interpretation, and recommending therapies based on the predicted stage of the disease. 

Although data-driven DL approaches have a huge potential to advance medical imaging technologies, the performance of these approaches rely heavily on the quality and availability of relatively large and annotated datasets for training.  However, obtaining such datasets in the medical domain is difficult and expensive. Medical image datasets are usually small or highly imbalanced due to several factors including patient privacy constraints, rare or complex diseases, and lack of diversity. They are also likely to be incompletely annotated due to tedious annotation. Further, datasets collected from real-world scenarios tend to contain missing and imperfect data due to sensor or background noise and corruptions. There might exist noisy labels due to ambiguity in annotations, human fatigue, and large inter/intra-user variations. Distribution/domain shifts which can significantly degrade the generalization performance of models are common and considerable for limited medical data given the variability of populations, disease manifestations, and imaging devices. Limited data may also lead to model bias, a particularly critical issue for medical applications in which DL decisions are of high consequence. To address training challenges in imperfect and data-limited scenarios, several techniques have been proposed, such as, weakly supervised learning, transfer learning, active learning, few shot learning, ensemble learning, and data augmentation, among others. Despite the promise of these techniques in a wide range of applications, there are still significant limitations and challenges when applied to medical images. There is also a lack of theoretical understanding and explanation of these techniques. Hence, new frameworks, algorithms, and techniques for coping with and learning from noisy data or data with limited annotations should be proposed to advance real-world modeling in medical image applications and improve the state-of-the-art of such research.