Mimic-Aware Tuberculosis Detection in Nigerian Chest X-rays via Disease-Specific Transfer Learning

Tuberculosis (TB) remains a leading cause of mortality in Nigeria, where chest X-ray (CXR) screening is complicated by radiographic overlap among TB, pneumonia, and COVID-19. Most AI models for TB detection rely on binary TB/normal classification, which does not evaluate differential diagnosis against these radiological mimics. We present a 4-class mimic-aware TB classification framework using DenseNet121, evaluating whether external TB-versus-normal pretraining (NIAID-CXR) improves differential diagnosis on a Nigerian CXR dataset (CXR-Kano). To address a confound in our original design, pretraining source and fine-tuning learning-rate strategy (layer-wise decay, LLRD, vs. uniform) varied simultaneously, we conducted a controlled 2×2 factorial ablation isolating each factor across four model configurations, evaluated via 5-fold cross-validation with soft-voting ensembling and paired statistical testing. TB-specific pretraining improved TB recall significantly when paired with LLRD (+5.3 points, bootstrap 95% CI [1.3, 9.3]) but showed no significant effect under uniform fine-tuning, revealing a genuine interaction rather than a main effect. A quantitative Grad-CAM analysis (fraction of activation energy within an anatomical lung mask) found that pretraining reduced lung-field attention concentration, consistently across both fine-tuning strategies (p=0.001, p=0.0001), the opposite of what a qualitative reading might suggest, and plausibly linked to a cross-source digitization confound we identified in the external pretraining dataset. These results show that pretraining's benefits are conditional on fine-tuning strategy and come with a measurable, previously undocumented cost to attention localization, underscoring the need for factorial ablation and quantitative interpretability audits, not qualitative comparison alone, in resource-constrained medical imaging research.