Development of AI-Assisted Decision Support Tools for Plastic Surgery

Abstract

Accurate localization of the Nipple-Areola Complex (NAC) is essential for achieving optimal aesthetic outcomes in chest reconstruction surgeries such as gender-affirming surgeries and post-mastectomy procedures. Traditional methods for NAC placement rely heavily on surgeons' experience and judgement, which are subject to variability and limited reproducibility. This thesis presents a machine learning-based decision support framework for automated NAC localization using pose estimation and regression modeling. A dataset of 102 clinical images from 34 male subjects was processed using OpenPose, an open-source pose estimation algorithm, to extract upper-body landmarks. Normalized distances between selected landmarks were used as input features for six regression models: Decision Tree Regressor, Random Forest Regressor, CatBoost Regressor, Multilayer Perceptron (MLP) Regressor, Support Vector Regressor (SVR), and Linear Regressor. Among these, the Linear Regression model achieved the lowest mean absolute percentage error (MAPE) of 0.06% for normalized distance prediction and 0.69% for average predicted NAC position. To further enhance accuracy, a tilt correction algorithm was implemented, reducing the MAPE from 1.2% to 0.75% (right NAC) and from 0.99% to 0.63% (left NAC) in cases with subject tilt.

To address limitations in dataset size and diversity, a second phase of research introduced a synthetic image generation pipeline using a Large Language Model (Llama 3.2). A prompt engineering strategy enabled the creation of demographically varied anatomically realistic images. Three experiments evaluated the impact of synthetic data on model performance. Results demonstrated that moderate augmentation improved performance and generalization, particularly when tested on demographically diverse or real-world images not seen during training.

This work illustrates the potential of AI-assisted decision support tools in plastic surgery, specifically in standardizing NAC localization through data-driven modeling. The integration of synthetic datasets further addresses challenges in data scarcity and enhances the generalizability of clinical machine learning applications.