DiffKD-DCIS: Predicting Upgrade of Ductal Carcinoma In Situ with Diffusion Augmentation and Knowledge Distillation

TL;DR

This work tackles DCIS upgrade prediction from ultrasound by addressing data scarcity with conditional diffusion-based data augmentation and improving generalization through a two-stage anatomy-aware knowledge distillation framework. It integrates a latent diffusion model guided by multimodal inputs (text prompts, tumor masks, and class labels) and a US-VAE to generate high-fidelity synthetic ultrasound images, then trains a high-capacity teacher and a compact student to transfer robust diagnostic reasoning. Across a multicenter dataset, DiffKD-DCIS achieves an external AUC of approximately 0.81, matches senior radiologist performance, and offers fast inference (0.15 s per case) with a small, deployment-friendly model. The study demonstrates a practical, expert-level AI tool for preoperative DCIS management and provides a blueprint for translating conditional generative-discriminative medical AI to real-world clinical practice.

Abstract

Accurately predicting the upgrade of ductal carcinoma in situ (DCIS) to invasive ductal carcinoma (IDC) is crucial for surgical planning. However, traditional deep learning methods face challenges due to limited ultrasound data and poor generalization ability. This study proposes the DiffKD-DCIS framework, integrating conditional diffusion modeling with teacher-student knowledge distillation. The framework operates in three stages: First, a conditional diffusion model generates high-fidelity ultrasound images using multimodal conditions for data augmentation. Then, a deep teacher network extracts robust features from both original and synthetic data. Finally, a compact student network learns from the teacher via knowledge distillation, balancing generalization and computational efficiency. Evaluated on a multi-center dataset of 1,435 cases, the synthetic images were of good quality. The student network had fewer parameters and faster inference. On external test sets, it outperformed partial combinations, and its accuracy was comparable to senior radiologists and superior to junior ones, showing significant clinical potential.

Figures (8)

• Figure 1: Overall architecture of DiffKD-DCIS. (A) Multi-conditional diffusion model guided by text prompts, tumor masks, and upgrade labels to generate high-fidelity synthetic ultrasound images. (B) Teacher network trained on real + synthetic augmented dataset. (C) Lightweight student network distilled from the teacher, maintaining comparable performance while outperforming all baselines on two independent external test sets.
• Figure 2: Architecture of the Conditional Latent Diffusion Model for Ultrasound Image Synthesis. A: Conditional Diffusion Backbone — iteratively denoises the latent representation $Z_t$ using the current time step t, noise level, and multi-modal conditioning signals (upgrade labels, tumor masks, and text prompts). Conditioning signals are injected via contact mechanisms in each U-Net block to precisely guide the generation of clinically relevant structures. B: Residual Block of U-Net — composed of Conv2d, LayerNorm, and skip connections. C: MidBlock of U-Net — employs GroupNorm, and SiLU activation to effectively model complex spatial and conditional relationships in the latent space. D: Multi-Layer Perceptron (MLP) — processes time-step embeddings and conditioning embeddings (labels $\&$ text) before feeding them into the U-Net. E: Ultrasound-Optimized Variational Autoencoder (VAE). Left: Encoder compresses raw ultrasound images into compact low-dimensional latent representations $Z_0$; Right: Decoder reconstructs high-fidelity ultrasound images from the final denoised latent codes $Z_0$.
• Figure 3: Architecture and training pipeline of the knowledge distillation framework for DCIS upgrade prediction. (A) Network architectures of the teacher and student models. Teacher model (top): A deeper CNN comprising four convolutional blocks (each: Conv2D--ReLU--MaxPool) followed by three fully connected (FC) layers, outputting class logits for upgrade and non-upgrade outcomes. Student model (bottom): A lightweight CNN with three convolutional blocks and two FC layers, designed for efficient inference while preserving discriminative capability. (B) Knowledge distillation training workflow. The pre-trained teacher generates softened probability distributions (soft labels) from input ultrasound images; these are combined with ground-truth hard labels to jointly supervise the student network via a hybrid loss.
• Figure 4: Qualitative Comparison of Synthetic Ultrasound Images Across Methods and Clinical Cases
• Figure 5: ROC curves and confusion matrices of different deep learning models evaluated on an identical external test set