According to reports presented on November 26 at RSNA 2023 in Chicago, the utilization of a deep-learning AI system designed to evaluate the cancerous nature of cardiac lesions smaller than 4 inches could potentially reduce the instances of both overtreatment and undertreatment of liver cancer.
As per insights shared by Aakanksha Rana, PhD, from Johnson & Johnson in New Brunswick, New Jersey, this AI engine functions by utilizing a preventive CT scan as a reference, pinpointing the tumor’s location through segmentation, and subsequently automating the assessment of the cancer likelihood within that area.
Rana mentioned that despite the promising outcomes, there is still considerable groundwork ahead, stating, “The results were promising considering the limited dataset available.” She also noted the significant impact of neglecting such minor abnormalities.
A substantial 42% of renal cell carcinomas (RCC), constituting approximately 80% of liver tumors, fall under the phase1A classification. Rana and her team embarked on this study to address the diagnostic challenges posed by these lesions due to their small size.
In order to train and validate the deep learning model, which operates in two phases—first, localizing the kidney and lesion regions using an nnUnet model, and then distinguishing between benign and malignant tumors using 3D Mobilnet models—data from 289 T1a patients with confirmed histopathology were employed. The accuracy of the model was verified against standard segmentations.
The AI system achieved an impressive AUC of 0.87 and an accuracy of 0.82 in tumor classification. Notably, it exhibited a sensitivity of 0.89 and a precision of 0.81 in detecting malignancies. Additionally, the model demonstrated a high level of precision with dice scores reaching 0.83 for precise localization, as highlighted by Rana.
This research underscores the potential to mitigate the diagnostic challenges associated with T1a RCCs given their prevalence. The final “localize to classify” framework of convolutional neural networks (CNNs) effectively segmented the tumors for accurate localization, distinguishing between benign and malignant T1a renal masses. However, further refinement and validation on a larger and more diverse dataset are essential before widespread application.
The deep learning model’s capability to precisely localize the area and provide an AI-based probability score for assessment could prove invaluable, especially in scenarios where a radiologist encounters uncertainties regarding small urinary masses. Rana emphasized the pivotal role of this technology in focusing on T1 urinary cases evident in CT images, describing it as a significant advancement in the field.
