Prostate cancer (PCa) is the second leading cause of cancer-related deaths in men worldwide, with the American Cancer Society projecting approximately 288,300 new cases in the United States in 2023 alone. Early and accurate detection is essential for improving survival outcomes.
Traditional diagnosis relies on the Gleason Score (GS), derived from biopsy samples analyzed by pathologists. However, this method suffers from significant interobserver variability (differences between pathologists) and intraobserver variability (inconsistencies by the same pathologist over time), making it prone to diagnostic errors.
Recent advances in multi-parametric MRI (mpMRI) have emerged as a powerful noninvasive tool for assessing prostate cancer risk. When combined with computer-aided diagnosis (CAD) systems driven by artificial intelligence, mpMRI has the potential to dramatically improve both the accuracy and consistency of prostate cancer grading.
Despite promising results from deep learning approaches, a key challenge remains: deep neural networks typically require large labeled datasets for training, which are scarce in medical imaging. Transfer learning offers a solution by borrowing knowledge from models trained on millions of natural images and applying it to medical scans with limited data.
The researchers developed a Multi-Parameter Transfer Learning (MPTL) model that simultaneously processes three types of MRI sequences: T2-weighted sagittal, T2-weighted transaxial, and Apparent Diffusion Coefficient (ADC) images. Each sequence captures different biological properties of the prostate, and combining them gives a richer, more complete picture than any single scan alone.
The model architecture uses three parallel branches, one for each MRI sequence type. Each branch applies the EfficientNet deep learning architecture, pre-trained on the large ImageNet dataset. The features extracted from each branch are then concatenated into a single combined feature vector, enabling the model to leverage complementary information from all three imaging modalities simultaneously.
Fine-tuning was applied to the pre-trained EfficientNet models by modifying the top layers and adding a Global Average Pooling (GAP) layer before the final classification step. This adaptation allows the model to focus on the most diagnostically relevant features while benefiting from the general visual knowledge already embedded in the pre-trained network.
The dataset consisted of 221 prostate cancer cases from the Cancer Imaging Archive, with 98 low-grade cases (Gleason Score 3+3, 3+4) and 123 high-grade cases (Gleason Score 4+3, 4+4, 5+3, 3+5). Data was split with 50% for training, 30% for validation, and 20% held out as a test set. Data augmentation techniques were applied to artificially expand the training set and reduce overfitting.
EfficientNet is a family of convolutional neural networks that use a technique called compound scaling to simultaneously balance network depth, width, and input resolution. This makes EfficientNet models significantly more efficient than older architectures -- they are reported to be 8.4 times smaller and 6.1 times faster than comparable CNNs without sacrificing accuracy.
The researchers evaluated three variants of EfficientNet -- B0, B5, and B7 -- to identify which scale performed best on the prostate cancer classification task. Each variant increases in complexity and capacity, with B7 being the largest and most capable of capturing fine-grained features in medical images.
Training was conducted using the Adam optimizer over 150 epochs with a learning rate range from 1e-1 to 1e-6 and weight decay of 1e-4. Experiments were run on a Google Colab Pro platform with a GPU (T4 P100) and 25 GB of memory. Input images were resized to 244 x 244 pixels to fit the EfficientNet input requirements.
A key aspect of the training strategy was using the pre-trained EfficientNet as a feature extractor by removing the final classification layer and treating all earlier layers as fixed. This approach dramatically reduces training time and avoids the risk of degrading useful pre-learned features, especially important when working with small medical datasets.
The best-performing model, EfficientNet-B7 with multi-parametric fusion (MPTL-B7), achieved an accuracy of 88.89%, precision of 91.67%, recall of 88.0%, and an F1-score of 89.47%. This significantly outperformed all baseline methods tested in the study.
In direct comparison with prior methods, the MPTL model surpassed VGG-16 (83% accuracy), ResNet (82.09%), InceptionV3 (80.09%), CART (82%), and GoogleNet (60.9%), demonstrating the clear advantage of the multi-parametric fusion approach over earlier single-architecture techniques.
Experiments comparing single-modality inputs showed that using only ADC images yielded 74.72% accuracy, while using only T2w sagittal reached 71.43%. Importantly, combining all three MRI sequences consistently outperformed any single or paired combination of sequences, validating the design principle of the three-branch architecture.
The model's ROC curves and learning curves demonstrated stable training with no signs of severe overfitting, and the performance gap between deep learning approaches and traditional machine learning methods confirmed that deep networks are more capable of capturing the complex visual patterns that distinguish low-grade from high-grade prostate cancer.
To assess whether the model could generalize beyond its training data, the researchers performed a cross-dataset validation experiment. The model was trained on 5,096 images from the Cancer Imaging Archive and then tested on 1,371 independent images from the I2CVB dataset, a completely separate data source with different image characteristics.
Despite differences in image shape, resolution, angle, and format between the two datasets, the MPTL model achieved an accuracy of 86.65%, precision of 83.36%, recall of 89.18%, and an F1-score of 86.13% on the unseen dataset. This demonstrates strong generalization capability across varied real-world data sources.
This cross-dataset performance is especially significant because medical imaging data is notoriously heterogeneous -- images from different hospitals or scanners often look quite different. A model that performs well across datasets is far more likely to be useful in clinical practice than one that only works on a single curated dataset.
Radiologists currently face significant challenges in interpreting prostate MRI images, particularly in distinguishing clinically significant (high-grade) cancer from indolent low-grade cancer. Misclassification in either direction has serious consequences: overtreatment of low-grade cancer causes unnecessary side effects, while undertreatment of high-grade cancer allows dangerous disease progression.
The MPTL model is designed to function as a computer-aided diagnostic (CAD) assistant, providing objective and reproducible risk assessments alongside radiologist review. Rather than replacing radiologists, the system is intended to reduce the burden of manual image interpretation and minimize the impact of observer variability on diagnostic outcomes.
The model's high precision (91.67% for B7) means that when it identifies a case as high-grade cancer, it is very likely to be correct -- an important property for clinical safety. By flagging suspicious cases automatically and consistently, the tool could help radiologists prioritize their attention and reduce the risk of missed or delayed diagnoses.
Importantly, the researchers demonstrated strong performance even with a limited training dataset of 221 cases, which is far smaller than what most deep learning models require. This suggests the MPTL approach could be deployed effectively at medical centers where large annotated datasets are not available.
This study introduced the Multi-Parameter Transfer Learning (MPTL) framework, a novel approach to prostate cancer classification using multi-parametric MRI. By fusing features from three MRI sequences through parallel EfficientNet branches, the model achieves state-of-the-art accuracy of 88.89% while requiring only a relatively small training dataset.
The work highlights the value of transfer learning from natural image datasets (like ImageNet) to medical imaging tasks. Even though ImageNet images are entirely different from MRI scans, the visual feature detectors learned from millions of natural images provide a strong foundation that can be fine-tuned for medical classification with far less labeled data.
Future work could explore expanding the model to classify cancer into more than two grades, incorporating additional MRI sequences or imaging biomarkers, and validating the system in prospective clinical trials. Ultimately, the goal is to build AI tools that can reliably support radiologists in daily clinical practice, reducing both diagnostic errors and patient harm.