Abstract
Convolutional Neural Networks (CNNs) can estimate anthropometric measurements from 2D images, yet the opacity of CNN models hinders deployment in health-adjacent settings. We study explainability for regression by adapting and comparing Gradient-weighted Regression Activation Mapping (Grad-RAM) variants to four child measurements (height, weight, waist circumference, head circumference) using a DenseNet-121 backbone and a landmark-guided cropping pipeline. To the best of our knowledge, this is the first paper to apply explainable AI (XAI) methods to body measurement estimation tasks, specifically of children. We evaluate five formulations: (i) Grad-RAM (signed), (ii) a Grad-CAM reference, (iii) GradRAM (|P|), (iv) Grad-RAMP (element-wise, | · | after summation), and(v) Element-wise Grad-RAMP (|·| before summation). Quantitatively, testset MAE reaches 2.80cm (height), 1.30kg (weight), 2.30cm (waist), and 1.08cm (head), matching or improving prior ARAN baselines. Qualitatively, all Grad-RAM/RAMP variants exhibit task-aware attributions (e.g., bodywide focus for height; abdominal emphasis for waist) while also revealing reliance on contextual cues. In a targeted ablation, masking background consistently degraded accuracy across tasks, confirming sensitivity to nonanatomical signals. We use these explanations not merely as post hoc audit tools but to position it in a cycle that feeds back to data collection and model training: saliency maps diagnose dataset and preprocessing biases (e.g., ruler dependence), guide data collection (background diversity), inform augmentation (cropping and aspect-ratio handling), and motivate model choices. This cycle operationalizes Grad-RAM for improving CNN regressors in imagebased body measurement estimation task.

This publication has 12 references indexed in Scilit: