FRAME: Faithful Radiologist–Audited Mechanistic Explainability for Ultrasound Imaging

Ultrasound imaging is widely used for diagnosis and monitoring because it is safe, portable, cost‐effective and suitable for real‐time clinical assessment. However, speckle noise, acoustic shadowing, attenuation, low tissue contrast and operator‐dependent variability remain challenging for automated prediction and spatial localisation. Although deep learning has improved ultrasound analysis, many models remain difficult to audit clinically. Common post hoc explanation methods, including Grad‐CAM, integrated gradients (IGs), LIME and SHAP, can produce diffuse, unstable or anatomically inconsistent attribution maps whose functional relationship with the underlying prediction is not guaranteed. We present FRAME, an explanation‐by‐design framework in which prediction and explanation are coupled through the same spatial evidence field. A lightweight U‐Net–style encoder–decoder generates dense evidence, and the scalar output is obtained directly by global average pooling of this field. Expert masks provide spatial regularisation during training. FRAME was evaluated on complete held‐out test sets across four ultrasound settings: BUSI breast ultrasound, microultrasound prostate, ulcerative colitis transabdominal ultrasound and FOCUS foetal four‐chamber ultrasound. BUSI was evaluated as diagnostic image‐level classification with spatial explanation, whereas the remaining datasets were treated as anatomy‐centred evidence‐localisation tasks. FRAME achieved Dice scores of 0.901 ± 0.061, 0.893 ± 0.048, 0.831 ± 0.031 and 0.887 ± 0.058 on BUSI, prostate, ulcerative colitis and FOCUS, respectively, with corresponding XAlign scores of 0.916 ± 0.052, 0.910 ± 0.043, 0.852 ± 0.028 and 0.905 ± 0.049. These values exceeded Grad‐CAM, IG, LIME and SHAP, as well as supervision‐controlled and mask‐free baselines. Empirical faithfulness analyses using deletion, insertion, sufficiency, comprehensiveness, evidence masking, model randomisation and perturbation robustness further showed that FRAME evidence was more strongly related to the model output. Component‐wise ablations confirmed contributions from evidence‐field supervision, global average pooling, focal loss, test‐time augmentation and the lightweight architecture. Independent radiologist and clinician assessment also showed high acceptance of FRAME explanations. Analysis of incorrect predictions demonstrated reduced spatial alignment, highlighting that FRAME remains susceptible to incomplete or off‐target evidence in failure cases. Overall, FRAME provides a computationally efficient and empirically validated approach for coupling prediction with anatomically grounded evidence in heterogeneous ultrasound applications.