PPubMed20 Oct 2019Cited 25×

Cascade and Fusion of Multitask Convolutional Neural Networks for Detection of Thyroid Nodules in Contrast-Enhanced CT

Zuopeng Zhao, Chen Ye, Yanjun Hu, Ceng Li, Xiaofeng Li

Abstract

With the development of computed tomography (CT), the contrast-enhanced CT scan is widely used in the diagnosis of thyroid nodules. However, due to the artifacts and high complexity of thyroid CT images, traditional machine learning has difficulty in detecting thyroid nodules in contrast-enhanced CT. A fully automated detection algorithm for thyroid nodules using contrast-enhanced CT images is developed. A modified U-Net architecture of fully convolutional networks is employed to segment the thyroid region of interest (ROI), and a fusion of convolutional neural networks (CNN-Fs) is proposed to detect benign and malignant thyroid nodules from the ROI images and original contrast-enhanced CT images. Experimental results demonstrate that the proposed cascade and fusion method of multitask convolutional neural networks (CNNs) is efficient in diagnosing thyroid diseases with contrast-enhanced CT images and has superior performance compared with other CNN methods.

A figure from Cascade and Fusion of Multitask Convolutional Neural Networks for Detection of Thyroid Nodules in Contrast-Enhanced CT — fig. from the paper

§ The Valyu brief

The paper, in five minutes

Read from the full text via Valyu Contents

A cascade and fusion of multitask CNNs, including DenseU-Net and CNN-F, achieves 95.73% accuracy for thyroid nodule detection in contrast‑enhanced CT.

What’s new

i.Proposed DenseU-Net, an improved U-Net architecture for segmenting thyroid contrast‑enhanced CT images.
ii.Developed a detection method that fuses multiple feature levels using transfer learning and two different CNN structures to classify benign and malignant nodules.
iii.Achieved strong performance without complex preprocessing or user intervention.

How they did it

The study first segmented the thyroid region using DenseU-Net, then trained two CNNs (CNN‑1 and CNN‑2) on ROI and original CT images, fusing them into CNN‑F. Data augmentation and five‑fold cross‑validation were applied, with transfer learning from ImageNet for CNN‑1 and Adam/SGD optimizers. Performance was evaluated using Dice coefficient for segmentation and accuracy, recall, precision, specificity, F1 score, and AUC for classification.

What they found

CNN‑F achieved 95.73% accuracy, 87.14% recall, 98.10% precision, 99.30% specificity, 92.30% F1 score, and 98.49 AUC (CI [97.91 99.06]).
DenseU-Net Dice coefficient: 0.987 (training) and 0.955 (validation); U‑Net Dice: 0.978 (training) and 0.945 (validation).
Validation set 1 (ROI images) average accuracy: 85.10%; validation set 2 (thyroid removed) average accuracy: 62.38%.

Caveats

✕DenseU-Net struggles with segmentation when background is complex and lesions blend with surrounding tissue.
✕CNN‑F cannot accurately classify cases at the benign‑malignant junction.
✕CNN hyperparameters were chosen empirically rather than analytically.
✕Performance may improve with larger, more diverse datasets.

§ The Valyu brief

Reading the full paper and taking notes. This takes a few seconds…

§ Ask this paper

Ask a question about this paper

Valyu reads the full text and answers from what the paper actually says.

§ Related work in other archives

Cascade and Fusion of Multitask Convolutional Neural Networks for Detection of Thyroid Nodules in Contrast-Enhanced CT

The paper, in five minutes

Ask a question about this paper

Similar papers in all four archives

Similar papers in all four archives

The paper, in five minutes