DiffULD: Diffusive Universal Lesion Detection

Peiang Zhao, Han Li, Ruiyang Jin, S. Kevin Zhou

TL;DR

This paper introduces DiffULD, a novel approach for Universal Lesion Detection (ULD) in CT scans that leverages Diffusion Probabilistic Models (DPMs). To overcome the challenge of insufficient training data in ULD, DiffULD employs a unique "center-aligned BBox padding" strategy to generate high-quality synthetic training targets, significantly improving detection performance.

What Problem Does This Paper Address?

Universal Lesion Detection (ULD) in medical imaging, particularly CT scans, is crucial for computer-aided diagnosis. The goal is to identify any lesion, regardless of its type or organ, which is a more challenging task than detecting specific types of lesions in a particular organ. Existing ULD methods often rely on anchor-based detection frameworks. However, these methods suffer from two main drawbacks:

  1. Insufficient Training Targets: Medical scans, especially CT slices, often contain very few lesions. This data imbalance makes it difficult for anchor-based methods to learn effectively, as there aren't enough positive anchors (regions likely to contain a lesion) to train on. Simply lowering the anchor-matching threshold leads to more false positives.
  2. Difficulties in Anchor Design: Lesions in CT scans vary greatly in size and shape, making it challenging to design a set of fixed anchor boxes that can effectively capture all of them. The rectangular nature of anchors can also be a poor fit for irregularly shaped lesions.

While anchor-free methods and Diffusion Probabilistic Models (DPMs) have shown success in natural image object detection, they struggle with ULD due to the lack of sufficient training targets and the absence of the "positional prior" that anchors provide.

Methods

DiffULD tackles these challenges by formulating lesion detection as a denoising diffusion process, similar to recent DPM-based object detectors like DiffusionDet. However, the core novelty lies in their proposed "Center-aligned BBox padding" strategy to address the insufficient training target problem.

Here's how it works:

  • Perturbated GT Generation: For each ground-truth (GT) bounding box (BBox), they generate multiple "perturbated" boxes. This involves two steps:
  • Box Scaling: The width and height of the GT BBox are randomly scaled within a defined range (controlled by Ascale).
  • Center Sampling: The center coordinates of the perturbated boxes are sampled from a 2D Gaussian distribution. Crucially, this sampling is done in a way that the centers of the perturbated boxes remain aligned with the center of the original GT BBox.
  • Perturbated GT Boxes: These perturbated boxes are then paired with their corresponding original GT BBoxes to form "perturbated GT boxes."
  • Forward Diffusion Process: These perturbated GT boxes are then corrupted with Gaussian noise over several steps (T steps), creating noisy versions (z_t) of the training targets. The model is trained to reverse this process, learning to denoise the noisy boxes back to the original perturbated GT boxes.
  • Reverse Diffusion Process (Inference): During inference, the model starts with a set of random boxes and iteratively refines them through the denoising process to predict the final lesion bounding boxes.
  • Box Filtering and Padding: To improve inference efficiency and robustness, they incorporate a box filtering step to remove low-confidence predictions and a crucial step of padding the set of remaining boxes with new random boxes to maintain a fixed number (N) for subsequent refinement steps. This prevents the model from running out of proposals too quickly.
  • Backbone: The model utilizes a multi-window input strategy, processing CT scans with three organ-specific HU windows to highlight different anatomical regions. These features are then fused using a modified 3D DenseNet backbone for context fusion.

Datasets

  • DeepLesion: The primary dataset used for training and evaluation. It's a large-scale dataset containing over 32,000 lesions from over 10,000 CT studies. The paper uses the official data split (70% train, 15% validation, 15% test).
  • Augmented DeepLesion Test Set: An additional augmented test set introduced by Lesion-Harvester [45] is used for further evaluation.

Key Insights and Ideas

  • DPMs are powerful but data-hungry: The paper highlights that while DPMs excel at natural image detection, their success relies on abundant training data. Applying them directly to ULD, with its sparse targets, leads to performance degradation.
  • Synthetic data quality matters: Simply adding random noise or boxes isn't enough. The "center-aligned BBox padding" strategy is key because it generates high-quality synthetic targets that are still closely related to the original lesions, effectively expanding the training set without introducing too much noise or irrelevant information.
  • Anchors provide a useful prior: The paper acknowledges that while they are moving away from traditional anchors, the idea of a positional prior is beneficial. Their padding strategy effectively mimics this by keeping the centers aligned, providing a structured starting point for the diffusion process.
  • Arbitrary box prediction is an advantage: Unlike fixed anchors, DPMs can predict boxes of arbitrary shapes and sizes, which is naturally suited for the diverse and irregular nature of medical lesions.

Key Results

  • State-of-the-art performance: DiffULD achieves competitive performance compared to several state-of-the-art ULD methods on the DeepLesion dataset, outperforming prior methods like DKA-ULD and SATr.
  • Significant improvement from padding: An ablation study shows that the "center-aligned BBox padding" strategy provides a substantial improvement (1.13%) over a baseline without padding, and outperforms other padding strategies like duplicating GT boxes or padding with random boxes.
  • Competitive on augmented data: The method also demonstrates strong results on the augmented DeepLesion test set, achieving higher sensitivity than established methods like A3D and SATr.

Limitations and Future Work

  • Computational Cost: Diffusion models are generally computationally intensive, and while the paper doesn't explicitly detail inference times, this is a common limitation.
  • Hyperparameter Tuning: The performance is sensitive to hyperparameters like Ascale, Aconf, and N (number of proposals).
  • Future Work: The paper suggests that the proposed method is promising for ULD and could be extended to other medical imaging tasks. The inherent ability of DPMs to handle diverse shapes and sizes makes them a good candidate for future research in medical image analysis.

Summary updated

Abstract

Universal Lesion Detection (ULD) in computed tomography (CT) plays an essential role in computer-aided diagnosis. Promising ULD results have been reported by anchor-based detection designs, but they have inherent drawbacks due to the use of anchors: (i) Insufficient training target and (ii) Difficulties in anchor design. Diffusion probability models (DPM) have demonstrated outstanding capabilities in many vision tasks. Many DPM-based approaches achieve great success in natural image object detection without using anchors. But they are still ineffective for ULD due to the insufficient training targets. In this paper, we propose a novel ULD method, DiffULD, which utilizes DPM for lesion detection. To tackle the negative effect triggered by insufficient targets, we introduce a novel Center-aligned bounding box (BBox) padding strategy that provides additional high-quality training targets yet avoids significant performance deterioration. DiffULD is inherently advanced in locating lesions with diverse sizes and shapes since it can predict with arbitrary boxes. Experiments on the benchmark dataset DeepLesion [32] show the superiority of DiffULD when compared to state-of-the-art ULD approaches.
BibTeX
@inproceedings{zhao2023diffuld,
  author = {Peiang Zhao and Han Li and Ruiyang Jin and S. Kevin Zhou},
  title = {{DiffULD: Diffusive Universal Lesion Detection}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  doi = {10.1007/978-3-031-43904-9\_10},
}
Figure