Attention-Augmented Deep Learning and Decision-Threshold Tuning for Primary Bone Tumor Classification in Radiographs

Primary bone tumors pose a significant oncological
problem wherein timely radiographic diagnosis is crucial. Auto-
mated learning algorithms usually encounter challenges of class
imbalance and overfitting because of small sample size of the
medical datasets. In this paper, we introduce a unifying approach
which uses an attention-based CNN (EfficientNetB4 with CBAM),
focal loss with batch sampling for handling class imbalance and
validation set-based threshold tuning for overfitting mitigation.
We evaluate the proposed approach on the BTXRD dataset and
conduct an investigation of methods which prevent overfitting
such as data augmentation, dropout, weight decay, learning
rate scheduling, early stopping, and label smoothing across five
neural networks (DenseNet121, EfficientNetB3, EfficientNetB4,
MobileNetV3-Large, Swin Transformer-Base) and a soft voting
ensemble. It turns out that default decision threshold (0.50)
produces low precision of the tumor class due to the high test set
imbalance (∼10:1). However, using validation set-based threshold
tuning (θ∗ = 0.8494), it is possible to increase precision from
0.43 to 0.85 while retaining 0.76 recall which results in 96.62%
accuracy, weighted F1-score of 0.9654, specificity of 0.9863, and
AUC-ROC of 0.9760.