SAM-Sode: Towards Faithful Explanations for Tiny Bacteria Detection

Explainability is critical for medical AI: The paper demonstrates that when AI systems help doctors detect bacteria in microscopy images, doctors need to understand why the AI made its decision—not just trust the result. This "confidence support" is essential for clinical adoption and patient safety.

Tiny objects pose unique explanation challenges: Traditional explanation methods fail for bacteria detection because the targets are extremely small with sparse features and are often obscured by complex backgrounds, resulting in blurry, unclear explanations that don't match what human experts actually see under the microscope.

Foundation models can improve AI transparency: The SAM-Sode framework uses a large pre-trained foundation model (SAM3) to refine and clarify explanations by adding geometric awareness, demonstrating that leveraging existing AI models' knowledge can make other AI systems more interpretable and trustworthy.

Dual constraints enhance explanation quality: By applying both "physical significance" and "geometric alignment" checks, the method reduces background noise in explanations while ensuring they align with actual bacterial morphology—showing that multiple validation layers produce explanations that better match expert intuition.

Responsible AI requires domain-specific solutions: The 40% improvement in explanation quality on their bacteria dataset (2,524 images) suggests that general-purpose explanation methods are insufficient; medical AI systems need specialized approaches tailored to their specific detection challenges.