Pulmonary Tuberculosis Detection Using an Ensemble of ConvNeXts

This study evaluates the effectiveness of a novel ensemble transfer learning (TL) method for tuberculosis (TB) detection. We employed the recently introduced ConvNeXt convolutional-based deep learning architecture as the backbone model. Two TL approaches were applied at intermediate layers of the pre-trained model (PTM), specifically ConvNeXt-Tiny, Small, and Base variants: Hard Cut, where features from an intermediate layer are directly passed to the classifier, and Soft Cut, where features from an intermediate layer are fine-tuned before classification. Experimental results on two widely used TB datasets-the Montgomery County (MC) and the Shenzhen (SZ) datasets demonstrate that the proposed ensemble TL method enhances detection performance. The Soft Cut approach yielded accuracy improvements of approximately 11.46% (Tiny), 8.71% (Small), 3.58% (Base) for MC, and 11.18% (Tiny), 10.27% (Small), 6.65% (Base) for SZ compared to the direct method, compared to the direct method.