Attention Guided Lymph Node Malignancy Prediction in Head and Neck Cancer

Purpose Accurate lymph node (LN) malignancy classification is essential for treatment target identification in head and neck cancer (HNC) radiation therapy. Given the constraints imposed by relatively small sample sizes in real-world medical applications, to classify LN malignancy status accurately, we proposed an attention-guided classification (AGC) scheme that (1) incorporates human knowledge (ie, LN contours) into model training to guide model’s “learning” direction, alleviating the critical requirement of large training samples by deep learning approaches; and (2) does not require accurate delineation of LNs in the inference stage but can highlight the discriminative region nearby the LN, which is important for malignancy determination. Methods and Materials In the proposed AGC scheme, there is an attention-guided convolutional neural network (agCNN) module, followed by a classification convolutional neural network (cCNN) module. The input of the proposed AGC scheme is a region of interest (ROI) containing the LN and its surrounding tissues. The agCNN is designed to find the discriminative region in the ROI, which outputs an activation map whose voxel values indicate the importance of the voxels in malignancy prediction. Through multiplying the activation map with the ROI, we obtain the input for the cCNN, which finally outputs the LN malignancy probability. To demonstrate the effectiveness of the proposed scheme, we performed experimental studies using positron emission tomography and contrast-enhanced computed tomography from 129 surgical HNC patients, including 791 LNs, with pathologic ground truth of malignancy status. To evaluate the performance, 5-folder cross validation was used. Results The sensitivity, specificity, accuracy, and area under the receiver operating characteristic (ROC) curve values obtained by the proposed AGC scheme were 0.91, 0.93, 0.92, and 0.98, respectively, significantly outperforming conventional convolutional neural network and radiomics approaches at a significance level of .05 under a paired ROC comparison statistical test. Conclusions We developed an AGC scheme that can highlight the discriminative region in an image for LN malignancy prediction, outperforming a conventional radiomics method that requires accurate segmentation and a standard convolutional neural network model without involving segmentation. Accurate lymph node (LN) malignancy classification is essential for treatment target identification in head and neck cancer (HNC) radiation therapy. Given the constraints imposed by relatively small sample sizes in real-world medical applications, to classify LN malignancy status accurately, we proposed an attention-guided classification (AGC) scheme that (1) incorporates human knowledge (ie, LN contours) into model training to guide model’s “learning” direction, alleviating the critical requirement of large training samples by deep learning approaches; and (2) does not require accurate delineation of LNs in the inference stage but can highlight the discriminative region nearby the LN, which is important for malignancy determination. In the proposed AGC scheme, there is an attention-guided convolutional neural network (agCNN) module, followed by a classification convolutional neural network (cCNN) module. The input of the proposed AGC scheme is a region of interest (ROI) containing the LN and its surrounding tissues. The agCNN is designed to find the discriminative region in the ROI, which outputs an activation map whose voxel values indicate the importance of the voxels in malignancy prediction. Through multiplying the activation map with the ROI, we obtain the input for the cCNN, which finally outputs the LN malignancy probability. To demonstrate the effectiveness of the proposed scheme, we performed experimental studies using positron emission tomography and contrast-enhanced computed tomography from 129 surgical HNC patients, including 791 LNs, with pathologic ground truth of malignancy status. To evaluate the performance, 5-folder cross validation was used. The sensitivity, specificity, accuracy, and area under the receiver operating characteristic (ROC) curve values obtained by the proposed AGC scheme were 0.91, 0.93, 0.92, and 0.98, respectively, significantly outperforming conventional convolutional neural network and radiomics approaches at a significance level of .05 under a paired ROC comparison statistical test. We developed an AGC scheme that can highlight the discriminative region in an image for LN malignancy prediction, outperforming a conventional radiomics method that requires accurate segmentation and a standard convolutional neural network model without involving segmentation.