Prediction of Ki-67 Status in Breast Cancer Using a Habitat-Guided 2.5D Multiparametric MRI Deep Learning Model.
TL;DR
The habitat-guided 2.5D deep learning model showed potential as a noninvasive imaging adjunct for preoperative Ki-67 status prediction in breast cancer and was rigorously evaluated against conventional 2D DL, clinical, and combined (DL+clinical) models using AUC, the DeLong test, and decision curve analysis (DCA).
Abstract
Objective
Intratumoral heterogeneity may limit the representativeness of biopsy-based Ki-67 assessment in breast cancer. We therefore developed and validated a habitat-guided 2.5D deep learning (DL) model based on multiparametric MRI for noninvasive preoperative prediction of high versus low Ki-67 expression.
Methods
This retrospective study enrolled 333 patients with invasive breast carcinoma from 2 distinct MRI vendor cohorts (Siemens, training set, n=233; United Imaging, independent test set, n=100). All patients underwent preoperative multiparametric MRI, including DCE-MRI and DWI. Hemodynamic parametric maps (wash-in, wash-out) and ADC maps were generated and subsequently clustered using a K-means algorithm (k=3) to create a functional habitat mask that quantitatively encodes intratumoral heterogeneity. A 7-channel 2.5D input tensor was then constructed by concatenating the central habitat-guided slice with its 6 adjacent anatomic slices. A ResNet18 backbone was trained to classify high (≥20%) versus low Ki-67 expression. The model's performance was rigorously evaluated against conventional 2D DL, clinical, and combined (DL+clinical) models using AUC, the DeLong test, and decision curve analysis (DCA).
Results
In the challenging independent cross-vendor test set, our habitat-guided DL25D model demonstrated superior performance, achieving an AUC of 0.821 (95% CI: 0.736-0.906) and a sensitivity of 0.804. It significantly outperformed both the conventional DL2D model (AUC: 0.654, P=0.002) and the clinical model (AUC: 0.686, P=0.019). The incorporation of clinical variables failed to yield further improvement (combined model AUC: 0.837, P=0.483 vs. DL25D; NRI=0.021, P>0.05). DCA confirmed the superior net clinical benefit of our approach across a wide spectrum of threshold probabilities. Importantly, Grad-CAM visualizations revealed that the habitat-guided model strategically focused its attention on intratumoral core regions, whereas the conventional 2D model was distracted by tumor margins and background tissue.
Conclusions
The habitat-guided 2.5D deep learning model showed potential as a noninvasive imaging adjunct for preoperative Ki-67 status prediction in breast cancer. Multicenter prospective validation is required before clinical use.