A deep learning model using hyperspectral image for EUS-FNA cytology diagnosis in pancreatic ductal adenocarcinoma

Cancer Medicine 2023 AI 5 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Page [1, 2]
The Challenge of Diagnosing Pancreatic Cancer Under a Microscope

Pancreatic ductal adenocarcinoma (PDAC) is one of the most deadly cancers, with incidence rising globally and a very poor prognosis. Accurate pathological diagnosis is essential for treatment planning, but it remains one of the most challenging tasks in cancer pathology.

The standard diagnostic approach is endoscopic ultrasound-guided fine needle aspiration (EUS-FNA), where a needle guided by ultrasound extracts cells directly from the pancreatic mass. While this procedure has 85-98% sensitivity and specificity, correctly reading the cytology slides can be difficult even for experienced pathologists.

PDAC cells can look deceptively similar to cells from chronic pancreatitis, an inflammatory condition of the pancreas. This overlap in appearance means that distinguishing cancer from benign inflammation requires expertise and time, and errors can have serious consequences for patients.

TL;DR: PDAC diagnosis from biopsy samples is challenging because cancer cells can look similar to benign inflammatory cells, motivating the development of AI-assisted diagnosis.
Page [2, 3]
Hyperspectral Imaging: Seeing Beyond What the Eye Can See

Standard microscopy captures images in three color channels (red, green, blue), but hyperspectral imaging (HSI) captures 40 separate spectral bands across the 450-750 nanometer wavelength range. This allows the camera to detect subtle chemical and structural differences in tissue that are invisible to the naked eye.

Each point in a hyperspectral image contains a full light spectrum, revealing information about the molecular composition of cells. Cancerous cells undergo chemical changes as they transform, and HSI can pick up spectral signatures of these changes that standard imaging simply cannot capture.

The researchers built a custom medical hyperspectral imaging system at Shanghai Key Laboratory, coupling an optical microscope with an acousto-optic tunable filter to capture 40-band spectral data from liquid-based cytology slides prepared from EUS-FNA specimens.

TL;DR: Hyperspectral imaging captures 40 wavelength bands instead of just 3 colors, revealing chemical and molecular differences between cancerous and benign cells that standard microscopy misses.
Pages 4-4
Training a Neural Network to Diagnose Pancreatic Cancer

The team collected pancreatic EUS-FNA specimens from 33 benign cases and 39 confirmed PDAC cases, generating 1,913 hyperspectral images total. Two experienced cytopathologists with over 10 years each reviewed and annotated regions of interest on the slides.

A ResNet18 convolutional neural network was combined with SimSiam self-supervised learning to build the classification model. The team used a multi-staged training approach under pathologist supervision, progressively refining the model's understanding of what makes a cancer cell distinct.

Attribution Guided Factorization Visualization (AGF-Visualization) was applied to make the model interpretable, highlighting which parts of each image were driving the diagnostic decision. This transparency is essential for building trust with clinicians who need to understand AI recommendations.

TL;DR: A ResNet18 neural network was trained on 1,913 hyperspectral images of pancreatic biopsies with pathologist supervision, using visualization tools to make its decisions transparent.
Pages 7-7
Impressive Diagnostic Accuracy on Cancer vs. Benign Cells

The ResNet18-SimSiam model trained on hyperspectral images achieved an accuracy of 92.04%, sensitivity of 93.10%, and specificity of 91.23% for distinguishing PDAC from benign pancreatic tissue. The area under the ROC curve was 0.9625, indicating excellent overall discriminative ability.

Critically, the model outperformed equivalent models trained on standard RGB color images, confirming that the additional spectral information from hyperspectral imaging genuinely improves diagnostic performance rather than just adding complexity.

AGF-Visualization confirmed that the model was making decisions based on tumor cell nuclei - the same features that experienced pathologists use - validating that the AI had learned medically meaningful patterns rather than spurious correlations.

TL;DR: The hyperspectral AI model achieved 92% accuracy, 93% sensitivity, and 91% specificity for PDAC diagnosis, outperforming standard color image-based models.
Page [9, 10]
Reducing the Burden on Cytopathologists

Cytopathology is a highly specialized and time-intensive discipline, and there is a global shortage of qualified cytopathologists. An AI tool that can pre-screen EUS-FNA slides and flag suspicious cases could dramatically reduce the workload on human experts while maintaining high diagnostic quality.

The model's ability to identify the relevant regions of interest on slides - essentially showing pathologists where to look - could accelerate review time, reduce inter-observer variability, and standardize diagnostic quality across institutions with varying levels of expertise.

The study demonstrates that hyperspectral AI is feasible even with relatively small training datasets (72 cases), suggesting it could be developed further with larger multi-center studies to achieve clinical-grade validation and eventual deployment.

TL;DR: The AI system could reduce pathologist workload by pre-screening biopsy slides and highlighting key diagnostic features, potentially standardizing PDAC diagnosis across different institutions.
Citation: Open Access, 2023. Available at: PMC10501295.