Patch-based Carcinoma Detection on Confocal Laser Endomicroscopy Images - A Cross-site Robustness Assessment

Marc Aubreville, Miguel Goncalves, Christian Knipfer, Nicolai Oetter, Tobias Würfl, Helmut Neumann, Florian Stelzle, Christopher Bohr, Andreas K. Maier

TL;DR

Researchers have demonstrated that deep learning models trained to detect oral cancer can successfully generalize to detect cancer in the vocal cords, despite differences in anatomy and clinical settings. By using a patch-based CNN approach, the system achieved nearly 90% accuracy in cross-site testing, moving us closer to reliable, real-time "optical biopsies" for head and neck cancers.

What Problem Does This Paper Address?

Head and Neck Squamous Cell Carcinoma (HNSCC) is a major global health burden, often diagnosed at late stages when treatment is less effective. The current "gold standard" for diagnosis is an invasive physical biopsy followed by histopathology. While accurate, biopsies are slow, carry risks of infection, and can miss the most aggressive parts of a tumor due to limited sampling.

Confocal Laser Endomicroscopy (CLE) offers a solution: it’s a non-invasive imaging technique that provides 1000x magnification, allowing doctors to see cellular structures in real-time—essentially an "optical biopsy." However, CLE images are notoriously difficult to interpret, requiring extensive specialist training. This paper investigates whether a deep learning algorithm can not only automate this detection but also remain robust when moved from one anatomical site (the oral cavity) to another (the vocal cords).

Methods

The authors employed a patch-based Convolutional Neural Network (CNN) approach:

  • Patch Extraction: Instead of processing a whole high-resolution image at once, the algorithm breaks the circular CLE field of view into 80x80 pixel patches.
  • Classification & Fusion: Each patch is classified independently by the CNN. The resulting probabilities are then "fused" to provide a single classification (cancerous vs. healthy) for the entire image.
  • Training Strategy: They used the Adam optimizer within the TensorFlow framework, training for 60 epochs.
  • Evaluation: To ensure the model wasn't just "memorizing" specific patients, they used Leave-One-Patient-Out (LOPO) cross-validation. This is a rigorous standard where the model is tested on a patient it has never seen during training.

Datasets

The study utilized two primary datasets acquired from different clinical departments:

  1. Oral Cavity (OC): 116 image sequences from 12 patients (7,894 total images). This included various sub-sites like the inner lip, hard palate, and alveolar ridge.
  2. Vocal Cords (VC): 47 image sequences from 5 patients (4,425 total images).

The OC dataset is considered more "diverse" because it covers multiple tissue types and was subject to more motion artifacts (as patients were often awake), whereas VC images were typically captured under general anesthesia.

Key Insights and Ideas

  • The "Superset" Advantage: A model trained on the oral cavity generalized remarkably well to the vocal cords (89.45% accuracy). The authors suggest this is because the oral cavity dataset acts as a "superset," containing a wider variety of noise, artifacts, and tissue structures that prepare the model for the cleaner vocal cord images.
  • Asymmetric Generalization: Interestingly, the reverse didn't work as well. A model trained only on vocal cords struggled when applied to the oral cavity (68.53% accuracy). This highlights that for medical AI, training on "messier," more diverse data is often better for robustness than training on "clean" data.
  • Anatomical Independence: The study provides strong evidence that the cellular features of squamous cell carcinoma are consistent enough across different head and neck sites that a single specialized model could potentially serve multiple disciplines (ENT, oral surgery, etc.).

Key Results

  • Cross-Site Accuracy: The model trained on Oral Cavity data achieved 89.45% accuracy and an AUC of 0.955 when tested on Vocal Cords.
  • State-of-the-Art Comparison: These results are comparable to (and in some cases exceed) previous models that were trained and tested on the same anatomical site.
  • Combined Performance: When the datasets were concatenated (OC+VC), the model reached its peak performance with 90.81% accuracy and an AUC of 0.970.
  • Expert Rivalry: The algorithmic accuracy (approx. 90%) is competitive with experienced CLE specialists, who typically score in the 82-92% range.

Limitations and Future Work

  • Surface Level Only: CLE has a low penetration depth. It can detect surface-level epithelial changes but might miss "buried" tumors spreading within the submucosa.
  • Artifact Handling: Currently, images with heavy artifacts were manually removed before training. The authors aim to automate this "quality control" step in future iterations.
  • Sample Size: While the image count is high (thousands), the patient count (17) is still relatively small. Larger, multi-center trials are needed to claim universal generalization.
  • Ethical Constraints: The "healthy" control images were taken from clinically normal tissue in cancer patients. Truly healthy tissue from non-cancer patients is rarely available for study due to the ethical concerns of performing endoscopy on healthy volunteers.

Summary updated

Abstract

Deep learning technologies such as convolutional neural networks (CNN) provide powerful methods for image recognition and have recently been employed in the field of automated carcinoma detection in confocal laser endomicroscopy (CLE) images. CLE is a (sub-)surface microscopic imaging technique that reaches magnifications of up to 1000x and is thus suitable for in vivo structural tissue analysis. In this work, we aim to evaluate the prospects of a priorly developed deep learning-based algorithm targeted at the identification of oral squamous cell carcinoma with regard to its generalization to further anatomic locations of squamous cell carcinomas in the area of head and neck. We applied the algorithm on images acquired from the vocal fold area of five patients with histologically verified squamous cell carcinoma and presumably healthy control images of the clinically normal contra-lateral vocal cord. We find that the network trained on the oral cavity data reaches an accurac
BibTeX
@inproceedings{aubreville2018patch,
  author = {Marc Aubreville and Miguel Goncalves and Christian Knipfer and Nicolai Oetter and Tobias Würfl and Helmut Neumann and Florian Stelzle and Christopher Bohr and Andreas K. Maier},
  title = {{Patch-based Carcinoma Detection on Confocal Laser Endomicroscopy Images - A Cross-site Robustness Assessment}},
  booktitle = {BIOIMAGING},
  year = {2018},
  url = {https://openreview.net/forum?id=SnatnFgzGr},
}

Loading related content...

Figure