Related Experiment Video
Updated: Jun 15, 2025
13:01
Industrialized, Artificial Intelligence-guided Laser Microdissection for Microscaled Proteomic Analysis of the Tumor Microenvironment
Published on: June 3, 2022
3.6K
Artificial Intelligence Recognition Model Using Liquid-Based Cytology Images to Discriminate Malignancy and
Ryota Tanaka1, Yukihiro Tsuboshita2, Mitsuaki Okodo3
1Department of Thoracic and Thyroid Surgery, Kyorin University, Tokyo, Japan.
Summary
A deep learning convolutional neural network (DCNN) model accurately classifies lung cancer cytology, distinguishing between normal, adenocarcinoma, and squamous cell carcinoma. This AI tool shows promise for aiding clinical cytopathological diagnosis.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Artificial intelligence (AI) image recognition is increasingly applied in clinical settings.
- Lung cancer diagnosis relies heavily on accurate cytopathology.
- Automated analysis of cytology images can improve diagnostic efficiency.
Purpose of the Study:
- To develop and evaluate a deep learning convolutional neural network (DCNN) model for automated classification of lung cancer cytology.
- To assess the model's performance in distinguishing between normal, adenocarcinoma, and squamous cell carcinoma.
- To explore the potential of AI in supporting clinical cytopathological diagnosis.
Main Methods:
- A dataset of 9,141 Papanicolaou-stained liquid-based cytology samples was created from normal, adenocarcinoma (ADC), and squamous cell carcinoma (SQCC) surgical specimens.
- Whole-slide imaging was used to scan 45 prepared slides.
- The Densenet-121 DCNN architecture was employed for classification tasks, including malignancy prediction (normal vs. malignant) and histological type differentiation (ADC vs. SQCC), utilizing AdamW optimizer and 5-fold cross-validation.
Main Results:
- The DCNN achieved high performance in malignancy prediction: patch-level accuracy of 0.94 (sensitivity 0.97, specificity 0.85) and case-level accuracy of 0.91 (sensitivity 0.92, specificity 0.88).
- For SQCC prediction, patch-level accuracy was 0.90 (sensitivity 0.86, specificity 0.91), and case-level accuracy was 0.78 (sensitivity 0.73, specificity 0.82).
- The model demonstrated robust performance in classifying lung cancer types and predicting malignancy.
Conclusions:
- The developed DCNN model exhibits excellent performance in predicting lung cancer malignancy and histological subtypes from cytology images.
- This AI-powered tool has the potential to assist clinicians in cytopathological diagnosis, potentially enhancing diagnostic accuracy and efficiency.
- Further integration into clinical workflows could reinforce training and improve diagnostic outcomes.
