Adaptive Binary Focal Loss: Enhancing Radiograph Image Classification With Balanced Specificity and Sensitivity

ABSTRACT Convolutional neural networks (CNN) are widely used to classify radiograph images. Musculoskeletal disorders (MSD) of the upper extremity (which comprises upper body parts such as the shoulder, elbow, wrist, and hand, allowing movement, strength and fine motor skills). However, their performance is often limited by class imbalance and the presence of hard samples. Although approaches like ensemble models, capsule networks and regularised CNNs in groups can address these issues, they require substantial computational resources. The adoption of loss function does not require additional computational overhead. Focal loss prioritises hard samples (samples that are not easy to classify); it simultaneously suppresses the gradients for easy samples, which affects learning. This can reduce accuracy and create an imbalance between sensitivity and specificity, which is an undesirable outcome in medical diagnostics. To overcome these limitations, adaptive binary focal loss (ABFL) is proposed here, which combines the strengths of binary cross‐entropy and focal loss to achieve balanced learning between easy and hard samples. A balance parameter, , is introduced to adaptively weigh the contributions of binary cross‐entropy and focal loss. This approach is further extended to multi‐class classification tasks through the proposed adaptive categorical focal loss (ACFL). In addition, a procedure is introduced to automatically tune the three key hyperparameters , and based on the characteristics of the dataset. This eliminates the need for manual intervention. ABFL and ACFL are compared with seven existing loss functions using DenseNet‐169 and Inception‐v3 on musculoskeletal radiograph images (MURA), a digital database for screening mammography (DDSM) and a garbage classification dataset. Compared to focal loss, Cohen's kappa score performance improved by 33.70% in ABFL on the MURA finger dataset. Similarly, ACFL achieved improvements of 58.07% and 20.23% on the DDSM and garbage datasets, respectively, while maintaining balanced sensitivity and specificity. These results show the robustness and effectiveness of both ABFL and ACFL in handling class imbalance and hard samples in CNN‐based classification.