Quantum-enhanced deep learning for elastographic image-based characterization of cutaneous and subcutaneous masses

Improving the diagnostic precision of skin and subcutaneous lesion assessments is the main goal of this study. Although elastography is an effective method for determining tissue stiffness, the high-dimensional, non-linear data these scans generate frequently present challenges for traditional deep learning models. To capture fine-grained stiffness patterns that conventional models may overlook, this study proposes and evaluates a hybrid quantum-classical deep learning architecture.

A retrospective dataset of 520 elastographic images with histological confirmation is used in the investigation. The proposed methodology uses a multi-stage hybrid pipeline, in which initial feature vectors are extracted from raw images using a traditional Convolutional Neural Network (CNN). A parameterized quantum circuit layer receives these vectors. This layer recognizes complex textural patterns by performing extensive transformations within a high-dimensional Hilbert space. A final fully connected layer processes the output to categorize lesions as either benign or malignant. We present a proof-of-concept simulation of combining a parameterized quantum layer with a classical CNN backbone. Although the current classical software emulation introduces a latency trade-off (124.5 ms vs. 8.4 ms per image), our result can serve as a baseline for analyzing low-parameter feature representations before running on physical QPU hardware.

Compared to fully classical baselines, the hybrid model showed a notable performance improvement. Important conclusions include 94.3% diagnostic accuracy, 93.8% sensitivity, and 94.7% specificity. From 0.921 (classical) to 0.963 (hybrid), the Area Under the ROC Curve (AUC) rose by 4.2%. Testing revealed that, especially in situations with unclear stiffness patterns, the quantum layer was the main source of increased predictive power.

Quantum circuit integration offers more "expressive power" than conventional neural networks. The model efficiently captures non-linear correlations in tissue density that are typically flattened or ignored by classical architectures, as it operates in a high-dimensional Hilbert space. This implies that combining quantum and classical techniques is especially well-suited to medical imaging, where minute changes in data have a clinically significant impact.

The study concludes that when it comes to describing cutaneous and subcutaneous masses, the hybrid classical-quantum model performs noticeably better than conventional deep learning techniques. These results indicate a promising advancement in medical AI that could provide physicians with more accurate and reliable tools for the early identification of cancer using elastographic data.