Predicting the Response of Locally Advanced Cervical Cancer to Chemoradiation Therapy: A Comparison of Hybrid and Deep Learning Radiomics Models

Abstract:

Accurately predicting the response of locally advanced cervical cancer (LACC) patients to chemoradiation therapy (CRT) is crucial for tailoring personalized treatment strategies. This study aimed to compare the performance of hybrid and deep learning radiomics models in predicting CRT response in LACC patients.

Introduction:

Radiomics, the extraction of quantitative features from medical images, has revolutionized treatment response prediction and prognosis in various cancers. Hybrid radiomics (HCR) models, combining handcrafted radiomics features with clinical factors, have shown promise in predicting CRT response in LACC. However, deep learning radiomics (DLR) models, utilizing deep neural networks for feature extraction and classification, have demonstrated superior performance in medical image analysis tasks.

Methods:

A retrospective study was conducted involving 119 LACC patients who underwent CRT. Magnetic resonance imaging (MRI) scans were acquired before CRT, and treatment response was assessed using the Response Evaluation Criteria in Solid Tumors (RECIST) criteria. HCR models were constructed by combining handcrafted radiomics features from MRI scans with clinical factors. DLR models were developed using a transfer learning approach, fine-tuning a pre-trained convolutional neural network (CNN) on the LACC MRI dataset. The models were evaluated using the area under the curve (AUC) of the receiver operating characteristic (ROC) curve and sensitivity.

Results:

DLR models generally outperformed HCR models in predicting CRT response, though the difference was not statistically significant. DLR models showed improved uncertainty estimates, suggesting better generalizability. Regarding sensitivity, DLR models excelled in identifying patients unlikely to achieve complete remission after CRT. Integrating clinical factors enhanced the HCR model’s performance. When using only MRI data, the DLR model marginally outperformed the HCR model (AUC; 0.721 vs. 0.597). However, with clinical factors integrated, the difference between HCR and DLR models was less pronounced, with the DLR model showing a higher AUC (0.782 vs. 0.676).

Conclusion:

Both HCR and DLR models showed potential in predicting CRT response in LACC patients. Combining clinical factors with imaging features in radiomics prediction models can improve performance. Further external validation is needed before clinical application.

Additional Key Points:

• DLR models demonstrated improved uncertainty estimates and sensitivity compared to HCR models.
• Combining clinical factors and imaging features in radiomics models enhanced performance.
• Transfer learning with pre-trained CNNs is widely used in medical image classification tasks.
• DLR models’ interpretability remains a challenge, requiring further research.
• External validation using a larger dataset is necessary for clinical applicability.

Call to Action:

Radiomics holds great promise in personalizing LACC treatment. If you’re a medical professional seeking to enhance your understanding of radiomics in oncology, consider enrolling in our comprehensive online course. Explore our dynamic curriculum, featuring expert instructors, interactive modules, and real-world case studies, to gain the knowledge and skills needed to leverage radiomics for improved patient care.