Explainable and Uncertainty Aware Deep Learning for Concealed Prohibited Item Detection in Nonintrusive X-ray Inspection

Technical Summary: Explainable and Uncertainty Aware Deep Learning for Concealed Prohibited Item Detection in Nonintrusive X-ray Inspection

Problem Statement
Prohibited-item detection in security X-ray imagery remains a critical challenge due to the complex nature of baggage contents. Objects are often small, overlapping, weakly contrasted, or deliberately concealed, leading to dense clutter and obscured boundaries. While recent deep learning benchmarks (SIXray, OPIXray, PIDray) have advanced the field, existing solutions often treat detection accuracy, concealment sensitivity, spatial explainability, and predictive reliability as isolated objectives. Furthermore, high-performing detectors frequently lack calibrated confidence estimates, and models trained on one dataset often fail to generalize to others due to significant domain shifts in scanner hardware, pseudocoloring, and acquisition environments. There is a need for a unified framework that enhances sensitivity to concealed items while providing quantitative evidence of spatial grounding and predictive reliability.

Methodology
The study proposes CAU-XInspect, a concealment-aware, explainable, and uncertainty-aware deep learning framework built upon a YOLO26 detector. The core innovation is the Concealment-Aware Feature Enhancement Module (CAFEM), inserted immediately before the detection head. CAFEM operates on multi-scale feature maps through three complementary branches:

1. Occlusion Branch: Uses depthwise and pointwise convolutions to preserve local structural cues visible even when objects are partially exposed.
2. Contrast Branch: Computes local residuals by subtracting average-pooled representations to emphasize weak boundaries and subtle intensity differences.
3. Context Branch: Employs dilated convolutions to capture broader receptive fields, integrating surrounding evidence to aid recognition of partially visible items.

These branches are fused and processed through a spatial attention mechanism. The framework is evaluated through a controlled B1–B6 ablation study:

• B1: YOLO26 baseline.
• B2: YOLO26 + CBAM (conventional attention).
• B3: YOLO26 + CAAM (concealment-aware attention).
• B4: YOLO26 + CAFEM (proposed multi-branch enhancement).
• B5: B4 + Mask Guidance (auxiliary mask prediction head).
• B6: B5 + Attention–Mask Consistency Learning (Smooth-L1 penalty to align attention and mask maps).

To address the sparsity of prohibited items, a training-only concealment proxy was developed using object size, mask fill ratio, and crowding (overlap) to weight the loss function, without using the official "Hard/Hidden" test labels during training. The framework also integrates Monte Carlo Dropout for uncertainty estimation and Temperature Scaling for calibration.

Key Contributions

1. Concealment-Aware Feature Enhancement: Introduction of CAFEM, which integrates occlusion, contrast, and contextual cues to refine features specifically for difficult X-ray cases.
2. Controlled Ablation Protocol: A systematic B1–B6 evaluation isolating the contributions of conventional attention, concealment-aware attention, CAFEM, mask guidance, and consistency learning.
3. Difficulty-Stratified Evaluation: Performance is reported separately on the PIDray Easy, Hard, and Hidden subsets to distinguish improvements on concealed items from aggregate metrics.
4. Quantitative Explainability: Spatial grounding is measured via Attention IoU and Mask IoU, comparing learned attention/mask regions directly against annotated object support regions, rather than relying solely on visual heatmaps.
5. Reliability Assessment: The study evaluates calibration (Expected Calibration Error, Brier score) and uncertainty (Monte Carlo Dropout AUC) to determine if the model can distinguish reliable from unreliable predictions.
6. External Robustness Check: A zero-shot evaluation on an independent SIXray-derived subset (831 images, 5 common classes) to expose cross-dataset domain shift without fine-tuning.

Experimental Results
Experiments were conducted on the PIDray benchmark (47,677 images, 12 classes).

• Detection Performance: The B5 configuration achieved the highest overall mAP@0.50:0.95 (0.5991) and precision (0.8305). The full B6 model achieved the highest overall recall (0.7111).
• Concealment Sensitivity: On the Hidden subset, B6 improved recall from 0.3550 (baseline) to 0.4918 (a ~38.5% relative improvement) and the Hidden F1-score from 0.4600 to 0.5478.
• Spatial Explainability: Localization improved progressively across the ablation sequence. B6 achieved an Attention IoU of 0.7578 and a Mask IoU of 0.7777, indicating strong alignment between learned features and annotated prohibited items.
• Reliability: The model demonstrated low calibration error (ECE 0.0095) and strong error discrimination using detector confidence (AUC 0.9894) and Monte Carlo dropout uncertainty (AUC 0.8873).
• Zero-Shot Generalization: When applied to the SIXray-derived test set without fine-tuning, performance collapsed significantly. mAP@0.50:0.95 dropped to 0.0408, and recall fell to 0.0954, demonstrating substantial cross-dataset domain shift.

Significance and Claims
The paper positions CAU-XInspect as a technical proof of concept rather than an operationally validated security system. The primary significance lies in demonstrating that concealment-aware feature enhancement can improve sensitivity and spatial grounding for difficult prohibited-item cases while providing complementary reliability information.

The author explicitly states that while the framework improves detection within the PIDray domain, the severe performance degradation on the SIXray-derived subset reveals that strong internal performance does not guarantee cross-scanner robustness. Consequently, the study concludes that broader cross-scanner validation, domain adaptation, and human-in-the-loop evaluation are prerequisites before real-world deployment. The work advocates for evaluating X-ray inspection as an integrated framework encompassing accuracy, concealment sensitivity, explainability, and uncertainty, rather than focusing solely on mean Average Precision.