Enhanced Cascade Prostate Cancer Classifier in mp-MRI Utilizing Recall Feedback Adaptive Loss and Prior Knowledge-Based Feature Extraction

TL;DR

This work addresses automated ISUP grading of clinically significant prostate cancer from mpMRI by tackling data imbalance and interpretability. It introduces a Recall Feedback Adaptive Loss (RFAloss) that uses dynamic accuracy $a$ and recall $r$ signals, a Prior Knowledge-Based Feature Extraction (F-E) aligned with PI-RADS, and an Enhanced Cascade Classification with confidence refinement to produce clinically interpretable outputs. On the PI-CAI PICAI_BIAS dataset, RFAloss substantially boosts recall, with an additional ~12.9% recall gain when combined with F-E while preserving accuracy, and the cascade strategy yields a balanced, diagonal recall pattern even for hard ISUP classes (e.g., ISUP 5 ~40% recall). Collectively, these contributions improve non-invasive mpMRI-based PCa diagnosis, offering interpretable, recall-sensitive guidance that can assist clinicians in decision-making and intervention planning.

Abstract

Prostate cancer is the second most common cancer in males worldwide, and mpMRI is commonly used for diagnosis. However, interpreting mpMRI is challenging and requires expertise from radiologists. This highlights the urgent need for automated grading in mpMRI. Existing studies lack integration of clinical prior information and suffer from uneven training sample distribution due to prevalence. Therefore, we propose a solution that incorporates prior knowledge, addresses the issue of uneven medical sample distribution, and maintains high interpretability in mpMRI. Firstly, we introduce Prior Knowledge-Based Feature Extraction, which mathematically models the PI-RADS criteria for prostate cancer as diagnostic information into model training. Secondly, we propose Adaptive Recall Feedback Loss to address the extremely imbalanced data problem. This method adjusts the training dynamically based on accuracy and recall in the validation set, resulting in high accuracy and recall simultaneously in the testing set.Thirdly, we design an Enhanced Cascade Prostate Cancer Classifier that classifies prostate cancer into different levels in an interpretable way, which refines the classification results and helps with clinical intervention. Our method is validated through experiments on the PI-CAI dataset and outperforms other methods with a more balanced result in both accuracy and recall rate.

Figures (6)

• Figure 1: The overall diagram of the proposed method. After preprocessing the T2W, DWI and ADC images, we model the reporting criteria of prostate cancer in mpMRI and design the F-E algorithm to extract features. These results are added as an additional channel for training. Three classifiers are trined to refine the results, and the RFAloss is used to guide the training. The lower part illustrates how the RFAloss works. The accuracy and recall serve as dynamic parameters fed back to the loss function. The hyperparameters $M$, $n_1$, and $n_2$ control the feedback intensity.
• Figure 2: (a) The first row are original images of T2WI, ADC and DWI, and the second row is corresponding preprocessed images. (b) Illustration of the F-E with DWI as an example. (c) The final F-E result was obtained by the weighted addition of (b). (d) The 2d and 3d presentation of (c). The images from left to right show the original picture added directly, the original picture added with weights, and the weighted addition after F-E. A comparison reveals that our feature map significantly enhances regions with high signal across all images, and increase the contrast between peak values and other values.
• Figure 3: This figure illustrates the mechanism of the Recall Feedback Adaptive loss function, which is controlled by three parameters. Specifically, $n_1$ and $n_2$ determine the feedback sensitivity for accuracy and recall. Together with $M$, they affect the value of the parameter $\mathcal{A}$. The difference between $\mathcal{A}$ and $1-acc$ would change to focus on positive samples, thereby changing the search baseline and causing fluctuations in the loss value. This will finally guide the loss function towards increasing recall.
• Figure 4: Figures \ref{['n1pic']}, \ref{['n2pic']}, and \ref{['Cpic']} show the training loss curves for hyperparameters $n_1$, $n_2$, and $M$. The curves have been smoothed using Gaussian smoothing. The fluctuation of $n_1$ and $n_2$ decreases as their values decrease, while the fluctuation of $M$ is higher at 0.7 and slightly lower at 0.3 compared to 0.5.
• Figure 5: Illustration of the loss function descent during training. In terms of convergence trend, the recall loss quickly converges but gets stuck in a local optimal solution; CE loss and focal loss converge straightforwardly and rapidly. Our proposed RFA loss exhibits significant fluctuations during descent. When feedback is masked in RFAloss (represented by the gray dashed line), it shows rapid convergence without feedback. Therefore, the dynamic feedback mechanism induces fluctuations in the loss function, which helps in searching for optimal parameters.