A deep learning model incorporating MRI radiomics and pathological data improved the prediction of bone metastases in patients with prostate cancer.
A model combining a deep learning algorithm, MRI data, and pathological features from tissue samples demonstrated success in predicting bone metastases in patients with prostate cancer, according to study findings published in The Journal of Cancer Research and Clinical Oncology.
Existing literature suggests that the development of bone metastases is associated with significantly higher mortality rates in patients with prostate cancer. While whole-body bone scans are typically used for the early detection of bone metastases, standard criteria for this practice is lacking, wrote Yun-Feng Zhang and colleagues. In addition, symptoms and clinical indicators, such as BMI and PSA levels, can indicate metastases foci, but the sensitivity and specificity of these indicators are poor.
“Early bone metastases in most patients with [prostate cancer] lack typical clinical manifestations; Consequently, the absence of efficient early diagnostic methods frequently results in a delay in the initiation of treatment,” the researchers wrote. “Therefore, it is necessary to explore and identify simpler and more effective indicators to predict the risk of bone metastases in [patients with prostate cancer]. This will help guide treatment in clinical practice.”
A Combination of Deep Learning, Radiomics, and Pathomics
The composite model incorporated deep learning technology, radiomics, and pathomics.
“Radiomics is centered around the proficient extraction of quantitative image features to precisely depict the areas affected by lesions effectively,” Zhang and team wrote. “Pathomics is an innovative approach that combines pathology, imaging, and computer science to understand disease processes.”
The study included 211 patients who were diagnosed with prostate cancer between January 2017 and February 2023. Among participants, 106 had bone metastases and 105 did not, as indicated by whole-body scans. The investigators randomly selected patients for the model training group (n=169) or the validation group (n=42).
A radiologist and urologist used three MRI sequences—T2 weighted imaging, diffusion weighted imaging, and apparent diffusion coefficient—to segment the region of interest. Tissue samples provided insight into the pathological features of the disease.
The researchers applied deep learning with ResNet50, which they noted “performs exceptionally well in image classification tasks and can handle larger and more complex datasets.”
“Deep learning features were extracted from both MRI and histopathological images and analyzed to demonstrate their relationship with bone metastasis in [patients with prostate cancer], independent of traditional clinical and pathological risk factors,” Zhang and colleagues wrote.
Successful Prediction of Bone Metastases
The researchers extracted 2,553 radiomics features, 3,379 deep learning features, and 2,048 pathomics features, using least absolute shrinkage and selection operator (LASSO) regression to identify 44 radiomics features, 23 deep learning features, and 13 pathomics features that had non-zero coefficients, suggesting bone metastases. Using decision curve analysis and calibration curves, they determined that their models provided clinical benefit.
The best prediction model based on deep transfer learning had an area under the curve (AUC) of 0.89 (95% CI, 0.799-0.989), while the best models based on either pathomics or radiomics showed AUCs of 0.85 (95% CI, 0.714-0.989) and 0.86 (95% CI, 0.735-0.979), respectively.
The composite model—which combined radiomics, deep transfer learning, and pathomics features—demonstrated the most effectiveness overall, with an AUC of 0.93 (95% CI, 0.854-1.000).
“The [decision curve analysis] for the nomograms showed a superior net clinical benefit for the combined models, providing valuable guidance for clinicians in formulating treatment strategies,” the team wrote.
The study was subject to several limitations, including small sample size, limited geographical range, and retrospective design. Nevertheless, Zhang and team reported that their results are comparable to those of similar studies. “Multimodal radiomics and pathomics serve as valuable predictors of the risk of bone metastases in patients with primary [prostate cancer],” they concluded. “This information may change the clinical management strategy.”
