Towards Accurate Acne Detection via Decoupled Sequential Detection Head

Xin Wei, Lei Zhang, Jianwei Zhang, Junyou Wang, Wenjie Liu, Jiaqi Li, Xian JiangarXiv 2023 · OpenReview

TL;DR

This paper introduces a novel "Decoupled Sequential Detection Head" (DSDH) for more accurate acne detection in images. DSDH improves upon existing methods by explicitly separating and sequencing the tasks of predicting an acne lesion's location (offset) and size (scaling), addressing inherent challenges like ambiguous boundaries and varied lesion dimensions. The authors also introduce a new, high-quality dataset, ACNE-DET, to facilitate research in this area.

What Problem Does This Paper Address?

Accurate detection and diagnosis of acne lesions are crucial for effective treatment and monitoring. However, existing automated methods struggle due to two main challenges:

  1. Ambiguous Boundaries: The edges of acne lesions are often unclear, making it difficult to precisely delineate their extent, even for human experts. This ambiguity directly impacts the accuracy of localization and size estimation.
  2. Arbitrary Dimensions: Acne lesions can vary dramatically in size, from tiny, almost invisible spots to large, prominent ones. This wide range of scales poses a significant challenge for detection models.

Furthermore, the scarcity of high-quality, well-annotated datasets specifically for acne detection has hindered progress in developing robust algorithms.

Methods

The core contribution of this paper is the Decoupled Sequential Detection Head (DSDH), designed to be integrated into existing two-stage object detection frameworks (like Faster R-CNN or Mask R-CNN). DSDH introduces two key mechanisms:

  • Task-Decouple Mechanism: The process of refining a proposed bounding box is broken down into two distinct sub-tasks: predicting the offset (change in position) and predicting the scaling (change in size). These two tasks are handled by separate, parallel branches within the detection head. This separation addresses the inherent incompatibility between predicting location and size, leading to improved accuracy for both.
  • Task-Sequence Mechanism: Instead of performing offset and scaling prediction simultaneously, DSDH executes them sequentially. First, the offset is predicted and applied to refine the bounding box. Then, using the refined bounding box, the scaling is predicted. This sequential approach allows the model to gain a more comprehensive understanding of the lesion's dimensions, especially after the initial localization has been improved.

The authors also explored a "cascade version" of DSDH, building upon the idea of cascaded detection heads to further refine predictions.

Datasets

  • ACNE-DET: A newly introduced, high-quality dataset specifically for acne detection. It features:
  • High-resolution facial images acquired with professional equipment.
  • Rigorous annotation by six dermatologists.
  • Ten fine-grained lesion categories (e.g., Comedo-C, Papule, Pustule, Nodule) plus an "Other" category.
  • Contains 276 facial images with 31,777 labeled lesion instances.
  • The dataset is split into training and testing sets, ensuring images from the same patient are not in both.
  • ACNE04: A public benchmark dataset for acne detection, used for comparison.

Key Insights and Ideas

  • Offset and Scaling are Fundamentally Different: The paper strongly argues that treating offset and scaling prediction as a single, unified task is suboptimal. Decoupling them into separate, specialized branches significantly improves performance.
  • Sequence Matters: The order in which offset and scaling are predicted can impact results. The "offset then scaling" sequence proved to be more effective, likely because improving the location first provides a better context for estimating the size.
  • Two-Stage Detectors Benefit Most: The authors found that their DSDH method, when applied to two-stage detectors, yielded superior results compared to one-stage detectors, particularly for detecting larger lesions (APL). This suggests that the iterative refinement process of two-stage methods is well-suited for the complexities of acne detection.
  • Ambiguous Boundaries are a Major Hurdle: The paper repeatedly emphasizes how the inherent ambiguity of acne lesion boundaries limits performance, even for human experts. Their method aims to mitigate this by improving the precision of localization and size estimation.

Key Results

  • State-of-the-Art Performance: DSDH, particularly when integrated with Mask R-CNN, achieved state-of-the-art results on both the new ACNE-DET dataset and the public ACNE04 dataset.
  • Significant Improvements: On ACNE-DET, Mask R-CNN with DSDH improved AP by 2.3%, APS by 1.6%, APM by 3.8%, and APL by 3.2% compared to the original Mask R-CNN. Similar improvements were observed when DSDH was applied to Faster R-CNN and Cascade R-CNN.
  • Ablation Studies: Experiments confirmed the effectiveness of both the task-decouple mechanism (gaining 1.5% AP) and the task-sequence mechanism (gaining 1.3% AP). Combining them yielded the best results.
  • Lesion Category Performance: Visualizations showed DSDH performing well across most lesion categories, sometimes even fitting lesion areas better than ground truth annotations.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that even with their improvements, the inherent ambiguity of acne lesion boundaries remains a significant challenge, limiting ultimate performance gains.
  • The "Other" category in ACNE-DET, which lumps together various less common lesion types, naturally shows lower AP.
  • Future Work:
    • The authors suggest exploring the capability of DSDH as a general-purpose detection head for other types of objects, not just acne lesions.
  • Further research could investigate more advanced ways to handle the ambiguous boundaries and diverse scales of acne.

Summary updated

Abstract

Accurate acne detection plays a crucial role in acquiring precise diagnosis and conducting proper therapy. However, the ambiguous boundaries and arbitrary dimensions of acne lesions severely limit the performance of existing methods. In this paper, we address these challenges via a novel Decoupled Sequential Detection Head (DSDH), which can be easily adopted by mainstream two-stage detectors. DSDH brings two simple but effective improvements to acne detection. Firstly, the offset and scaling tasks are explicitly introduced, and their incompatibility is settled by our task-decouple mechanism, which improves the capability of predicting the location and size of acne lesions. Second, we propose the task-sequence mechanism, and execute offset and scaling sequentially to gain a more comprehensive insight into the dimensions of acne lesions. In addition, we build a high-quality acne detection dataset named ACNE-DET to verify the effectiveness of DSDH. Experiments on ACNE-DET and the public benchmark ACNE04 show that our method outperforms the state-of-the-art methods by significant margins. Our code and dataset are publicly available at (temporarily anonymous).
BibTeX
@misc{wei2023towards,
  author = {Xin Wei and Lei Zhang and Jianwei Zhang and Junyou Wang and Wenjie Liu and Jiaqi Li and Xian Jiang},
  title = {{Towards Accurate Acne Detection via Decoupled Sequential Detection Head}},
  year = {2023},
  eprint = {2301.12219},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2301.12219},
}
Figure