A hybrid deep learning magnetic resonance imaging (MRI) model that facilitates preoperative prediction of tumor deposits could have a significant impact on the staging and treatment of patients with rectal cancer.
In a new retrospective study, recently published in Insights into Imaging, researchers compared single-channel, multi-channel and hybrid deep learning (DL) models as well as a clinical model in a cohort comprised of 137 rectal cancer patients with tumor deposits and 190 rectal cancer patients without tumor deposits.
According to the study, the single-channel DL was only based on tumor regions of interest (ROI) and the multi-channel DL model incorporated tumor and peri-tumoral ROIs. The researchers said the hybrid DL model combined the multi-channel DL model with selected clinical data.
Independent testing of the models showed that the hybrid DL model offered a 21 percent higher sensitivity rate (77 percent) in predicting tumor deposits than the multi-channel DL model and clinical models (56 percent for both).
The hybrid DL model demonstrated a 6 percent higher specificity rate in comparison to the clinical model (85 percent vs. 79 percent). Researchers also noted the hybrid DL model provided a 9 percent and 13 percent higher positive predictive value (PPV) in comparison to multi-DL and clinical models (79 percent vs. 70 percent and 66 percent).
The study authors suggest the inclusion of peri-tumoral and tumor imaging may facilitate tumor deposit detection in multi- and hybrid deep learning models for patients with rectal cancer. The above images provide a saliency map analysis for a 67-year-old patient with tumor deposits. (Images courtesy of Insights into Imaging.)
Given the significance of tumor deposits with disease progression in rectal cancer and the challenges with traditional imaging in classifying tumor deposits, the study authors emphasized that the prognostic findings with the hybrid DL model could represent a significant advance in preoperative assessment.
“If we can predict the presence of (tumor deposits) preoperatively, we can avoid overestimating patients’ prognoses and refraining from forgoing (neoadjuvant chemoradiotherapy), which has the potential to significantly enhance patient outcomes, and improve patients’ survival rates,” wrote study co-author Bin Song, M.D., Ph.D., the director of the Department of Radiology at the West China Hospital of Sichuan University in Chengdu, China, and colleagues.
Three Key Takeaways
- Hybrid DL model improves predictive accuracy. The study highlights that a hybrid deep learning (DL) magnetic resonance imaging (MRI) model, combining multi-channel DL with selected clinical data, demonstrates significantly improved sensitivity (21 percent higher) and specificity (6 percent higher) compared to both the multi-channel DL model and the clinical model. This suggests that incorporating both imaging data and relevant clinical information enhances the preoperative prediction of tumor deposits in rectal cancer patients.
- Clinical impact on staging and treatment. The use of the hybrid DL model could have a substantial impact on the staging and treatment of patients with rectal cancer. The increased sensitivity and specificity of the model contribute to more accurate preoperative assessments, potentially avoiding overestimation of prognoses and guiding decisions on neoadjuvant chemoradiotherapy. This improvement in patient stratification could lead to enhanced outcomes and increased survival rates for individuals with rectal cancer.
- Importance of peri-tumoral fat regions in DL models. The study suggests that including peri-tumoral fat regions in DL models enhances the predictive capacity for tumor deposits. The multi-channel DL model, which incorporated tumor and peri-tumoral regions of interest (ROIs), demonstrated higher area under the curve (AUC) and specificity compared to the single-channel DL model. The researchers propose that analyzing peri-tumoral adipose tissue may contribute to the detection of tumor deposits, providing valuable insights into the potential mechanisms of tumor progression in rectal cancer.
The researchers noted that the multi-channel DL model demonstrated a higher AUC than the single-channel DL model (73.8 percent vs. 67.6 percent) as well as increased specificity (83 percent vs. 70 percent) in independent testing. Accordingly, the study authors suggested the inclusion of peri-tumoral fat regions may enhance the predictive capacity for tumor deposits with DL models.
“The potential mechanisms may be … (tumor deposits) via neurotrophic extravascular migratory and perineural invasion, which could lead to radiographic changes around the tumor regions … so the analysis of peri-tumoral adipose tissue could contribute to the detection of (tumor deposits),” maintained Song and colleagues.
(Editor’s note: For related content, see “Study Finds Questionable Adherence to Guidelines for Rectal MRI,” “Seven Key Considerations with Rectal Cancer MRI” and “Longer PET Acquisition Time Increases FDG-Avid Lymph Node Detection in Rectal Cancer.”)
Beyond the inherent limitations of a single-center retrospective study, the researchers noted that T2-weighted MRI images were utilized to build the deep learning model whereas multi-MRI sequences could bolster model applicability with other datasets. A lack of automatic delineation of target regions may comprise the model’s reproducibility, according to the study authors. They also noted the potential for patient selection bias given the study’s focus on limiting inclusion to those with rectal cancer that was confirmed by surgical pathology.