With the advancement of treatment modalities in radiation therapy for cancer patients, outcomes have improved, but at the cost of increased treatment plan complexity and planning time. The accurate prediction of dose distributions would alleviate this issue by guiding clinical plan optimization to save time and maintain high quality plans. We have modified a convolutional deep network model, U-net (originally designed for segmentation purposes), for predicting dose from patient image contours of the planning target volume (PTV) and organs at risk (OAR). We show that, as an example, we are able to accurately predict the dose of intensity-modulated radiation therapy (IMRT) for prostate cancer patients, where the average Dice similarity coefficient is 0.91 when comparing the predicted vs. true isodose volumes between 0% and 100% of the prescription dose. The average value of the absolute differences in [max, mean] dose is found to be under 5% of the prescription dose, specifically for each structure is [1.80%, 1.03%](PTV), [1.94%, 4.22%](Bladder), [1.80%, 0.48%](Body), [3.87%, 1.79%](L Femoral Head), [5.07%, 2.55%](R Femoral Head), and [1.26%, 1.62%](Rectum) of the prescription dose. We thus managed to map a desired radiation dose distribution from a patient’s PTV and OAR contours. As an additional advantage, relatively little data was used in the techniques and models described in this paper.

Throughout the course of delivering a radiation therapy treatment, which may take several weeks, a patient’s anatomy may change drastically, and adaptive radiation therapy (ART) may be needed. Cone-beam computed tomography (CBCT), which is often available during the treatment process, can be used for both patient positioning and ART re-planning. However, due to the prominent amount of noise, artifacts, and inaccurate Hounsfield unit (HU) values, the dose calculation based on CBCT images could be inaccurate for treatment planning. One way to solve this problem is to convert CBCT images to more accurate synthesized CT (sCT) images. In this work, we have developed a cycle-consistent generative adversarial network framework (CycleGAN) to synthesize CT images from CBCT images. This model is capable of image-to-image translation using unpaired CT and CBCT images in an unsupervised learning setting. The sCT images generated from CBCT through this CycleGAN model are visually and quantitatively similar to real CT images with decreased mean absolute error (MAE) from 69.29 HU to 29.85 HU for head-and-neck (H&N) cancer patients. The dose distributions calculated on the sCT by CycleGAN demonstrated a higher accuracy than those on CBCT in a 3D gamma index analysis with increased gamma index pass rate from 86.92% to 96.26% under 1 mm/1% criteria, when using the deformed planning CT image (dpCT) as the reference. We also compared the CycleGAN model with other unsupervised learning methods, including deep convolutional generative adversarial networks (DCGAN) and progressive growing of GANs (PGGAN), and demonstrated that CycleGAN outperformed the other two models. A phantom study has been conducted to compare sCT with dpCT, and the increase of structural similarity index from 0.91 to 0.93 shows that CycleGAN performed better than DIR in terms of preserving anatomical accuracy.

Better understanding of the dose-toxicity relationship is critical for safe dose escalation to improve local control in late-stage cervical cancer radiotherapy. In this study, we introduced a convolutional neural network (CNN) model to analyze rectum dose distribution and predict rectum toxicity. Forty-two cervical cancer patients treated with combined external beam radiotherapy (EBRT) and brachytherapy (BT) were retrospectively collected, including twelve toxicity patients and thirty non-toxicity patients. We adopted a transfer learning strategy to overcome the limited patient data issue. A 16-layers CNN developed by the visual geometry group (VGG-16) of the University of Oxford was pre-trained on a large-scale natural image database, ImageNet, and fine-tuned with patient rectum surface dose maps (RSDMs), which were accumulated EBRT + BT doses on the unfolded rectum surface. We used the adaptive synthetic sampling approach and the data augmentation method to address the two challenges, data imbalance and data scarcity. The gradient-weighted class activation maps (Grad-CAM) were also generated to highlight the discriminative regions on the RSDM along with the prediction model. We compare different CNN coefficients fine-tuning strategies, and compare the predictive performance using the traditional dose volume parameters, e.g. D _0.1/1/2cc , and the texture features extracted from the RSDM. Satisfactory prediction performance was achieved with the proposed scheme, and we found that the mean Grad-CAM over the toxicity patient group has geometric consistence of distribution with the statistical analysis result, which indicates possible rectum toxicity location. The evaluation results have demonstrated the feasibility of building a CNN-based rectum dose-toxicity prediction model with transfer learning for cervical cancer radiotherapy.

Accurate and automatic brain metastases target delineation is a key step for efficient and effective stereotactic radiosurgery (SRS) treatment planning. In this work, we developed a deep learning convolutional neural network (CNN) algorithm for segmenting brain metastases on contrast-enhanced T1-weighted magnetic resonance imaging (MRI) datasets. We integrated the CNN-based algorithm into an automatic brain metastases segmentation workflow and validated on both Multimodal Brain Tumor Image Segmentation challenge (BRATS) data and clinical patients' data. Validation on BRATS data yielded average DICE coefficients (DCs) of 0.75±0.07 in the tumor core and 0.81±0.04 in the enhancing tumor, which outperformed most techniques in the 2015 BRATS challenge. Segmentation results of patient cases showed an average of DCs 0.67±0.03 and achieved an area under the receiver operating characteristic curve of 0.98±0.01. The developed automatic segmentation strategy surpasses current benchmark levels and offers a promising tool for SRS treatment planning for multiple brain metastases.

Partly due to the use of exhaustive-annotated data, deep networks have achieved impressive performance on medical image segmentation. Medical imaging data paired with noisy annotation are, however, ubiquitous, but little is known about the effect of noisy annotation on deep learning based medical image segmentation. We studied the effect of noisy annotation in the context of mandible segmentation from CT images. First, 202 images of head and neck cancer patients were collected from our clinical database, where the organs-at-risk were annotated by one of twelve planning dosimetrists. The mandibles were roughly annotated as the planning avoiding structure. Then, mandible labels were checked and corrected by a head and neck specialist to get the reference standard. At last, by varying the ratios of noisy labels in the training set, deep networks were trained and tested for mandible segmentation. The trained models were further tested on other two public datasets. Experimental results indicated that the network trained with noisy labels had worse segmentation than that trained with reference standard, and in general, fewer noisy labels led to better performance. When using 20% or less noisy cases for training, no significant difference was found on the segmentation results between the models trained by noisy or reference annotation. Cross-dataset validation results verified that the models trained with noisy data achieved competitive performance to that trained with reference standard. This study suggests that the involved network is robust to noisy annotation to some extent in mandible segmentation from CT images. It also highlights the importance of labeling quality in deep learning. In the future work, extra attention should be paid to how to utilize a small number of reference standard samples to improve the performance of deep learning with noisy annotation.

We propose a novel BIRADS-SSDL network that integrates clinically-approved breast lesion characteristics (BIRADS features) into task-oriented semi-supervised deep learning (SSDL) for accurate diagnosis of ultrasound (US) images with a small training dataset. Breast US images are converted to BIRADS-oriented feature maps (BFMs) using a distance-transformation coupled with a Gaussian filter. Then, the converted BFMs are used as the input of an SSDL network, which performs unsupervised stacked convolutional auto-encoder (SCAE) image reconstruction guided by lesion classification. This integrated multi-task learning allows SCAE to extract image features with the constraints from the lesion classification task, while the lesion classification is achieved by utilizing the SCAE encoder features with a convolutional network. We trained the BIRADS-SSDL network with an alternative learning strategy by balancing the reconstruction error and classification label prediction error. To demonstrate the effectiveness of our approach, we evaluated it using two breast US image datasets. We compared the performance of the BIRADS-SSDL network with conventional SCAE and SSDL methods that use the original images as inputs, as well as with an SCAE that use BFMs as inputs. The experimental results on two breast US datasets show that BIRADS-SSDL ranked the best among the four networks, with a classification accuracy of around 94.23 ± 3.33% and 84.38 ± 3.11% on two datasets. In the case of experiments across two datasets collected from two different institutions/and US devices, the developed BIRADS-SSDL is generalizable across the different US devices and institutions without overfitting to a single dataset and achieved satisfactory results. Furthermore, we investigate the performance of the proposed method by varying the model training strategies, lesion boundary accuracy, and Gaussian filter parameters. The experimental results showed that a pre-training strategy can help to speed up model convergence during training but with no improvement of the classification accuracy on the testing dataset. The classification accuracy decreases as the segmentation accuracy decreases. The proposed BIRADS-SSDL achieves the best results among the compared methods in each case and has the capacity to deal with multiple different datasets under one model. Compared with state-of-the-art methods, BIRADS-SSDL could be promising for effective breast US computer-aided diagnosis using small datasets.