Treating dataset imbalance in fetal echocardiography classification

Deep learning has been a trending topic during the last few years, notably in medical imaging that employs neural networks for image manipulation, computer-aided detection of diseases, and many other tasks depending on the clinical practices. One possible application that would benefit from these methods is the fetal cardiac view classification, where these different views are useful to obtain valuable information about the patient's heart development. A trained network could help reduce variance in interpretation and speed up data annotation. Alas, in this context we can face two challenges: datasets may contain a lot of information not relevant to the outcome of the classifier's training, and the view classes may be unbalanced in the sense that certain classes may have much more samples than others. This paper presents a series of attempts to solve these issues and can be used as a practical guide for training viable classifiers in this context.

---

Opracowanie rekordu ze środków MEiN, umowa nr SONP/SP/546092/2022 w ramach programu "Społeczna odpowiedzialność nauki" - moduł: Popularyzacja nauki i promocja sportu (2022-2023).