Evaluating the Impact of Layer Depth and Data Availability on Multi-Output Deep Learning Models

Accurate age and gender estimation from skeletal radiographs is vital in medical diagnostics, enhancing patient management and treatment planning. This study evaluates the impact of layer depth and data availability on multi-output deep learning models, comparing a custom Convolutional Neural Network (CNN) and a custom DenseNet121. The dataset consists of 9,343 skeletal radiographs from healthcare facilities in Kurdistan, Iraq, encompassing diverse skeletal parts, age groups, and genders. Extensive preprocessing, including grayscale conversion, normalization, and augmentation techniques, was conducted to improve model robustness. The custom CNN, designed with six convolutional layers, incorporates batch normalization, dropout layers, and ReLU activation functions. DenseNet121 features dense connectivity, connecting each layer directly to facilitate efficient feature reuse and reduce parameter count. Both models were trained using the Adam optimizer with adaptive learning rates, employing checkpointing and early stopping. Evaluation metrics included Mean Absolute Error (MAE) Root Mean Square Error (RMSE) Mean Squared Error (MSE) for age estimation and accuracy, precision, recall, and F1-score for gender classification. Results show that the custom CNN achieved an MAE of 4.25 in age prediction, outperforming DenseNet121's MAE of 5.1. However, DenseNet121 excelled in gender classification with 92% accuracy compared to CNN's 90%.