Development and validation of an interpretable machine learning model for diagnosing pathologic complete response in...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Oncology
• Radiomics and Deep Learning

Background:

• Pathologic complete response (pCR) after neoadjuvant chemotherapy (NACT) is crucial for breast cancer prognosis, potentially enabling surgery omission.
• Accurate, noninvasive pCR diagnosis is challenging, especially when tumors disappear post-NACT, due to current imaging limitations.

Purpose of the Study:

• To develop and validate an innovative framework for noninvasive diagnosis of pCR in breast cancer patients.
• To address the challenge of analyzing imaging data where target lesions are absent post-NACT.

Main Methods:

• Developed a novel framework using Dimensional Accumulation for Layered Images (DALI) and an Attention-Box tool.
• Transformed 3D MRI to 2D, employed tissue-region normalization and subtraction imaging for preprocessing.
• Extracted radiomic and deep learning features, integrated into a stacked ensemble machine learning model.

Main Results:

• The stacked ensemble model achieved an AUC of 0.831 on the test set, outperforming individual models.
• Preprocessing techniques significantly improved diagnostic accuracy.
• SHAP analysis ensured model interpretability by identifying key predictive variables.

Conclusions:

• The innovative framework successfully addresses challenges in noninvasive pCR diagnosis.
• Advanced preprocessing and machine learning improve feature quality and model performance for surgical decision-making.
• This approach supports clinicians in identifying patients eligible for surgery omission, reducing overtreatment and enhancing quality of life.