Computed tomography-based 3D convolutional neural network deep learning model for predicting micropapillary or solid...

Area of Science:

• Radiology
• Oncology
• Artificial Intelligence

Background:

• Invasive lung adenocarcinoma (ILADC) is a major cause of cancer mortality.
• Accurate prediction of growth patterns, such as micropapillary or solid (M/S), is crucial for ILADC treatment and prognosis.
• Current methods for determining growth patterns often require invasive procedures.

Purpose of the Study:

• To evaluate the efficacy of a computed tomography (CT)-based deep learning (DL) model in predicting the M/S growth pattern of ILADC.
• To assess the model's performance in distinguishing ILADC subtypes noninvasively.

Main Methods:

• Two DL models were developed using a self-paced learning (SPL) 3D Net architecture on preoperative CT scans from 617 ILADC patients (training/internal validation) and 353 external validation patients.
• Model 1 predicted M/S pattern in all ILADC cases, while Model 2 focused on ILADC with a diameter ≤ 2 cm.
• Performance was evaluated using metrics including AUC, accuracy, recall, precision, and F1-score.

Main Results:

• Model 1 achieved an AUC of 0.857 and accuracy of 0.805 in the external validation set.
• Model 2 demonstrated strong performance with an AUC of 0.831 and accuracy of 0.792 in the external validation set.
• The SPL 3D Net model outperformed other DL architectures like ResNet and DenseNet.

Conclusions:

• A CT-based DL model can reliably predict M/S growth patterns in ILADC.
• This AI tool offers a noninvasive approach for detecting and differentiating ILADC subtypes.
• The model is effective even for small-sized ILADC tumors, aiding in early diagnosis and treatment planning.