Medical image synthesis via conditional GANs: Application to segmenting brain tumours

Accurate brain tumour segmentation is critical for tasks such as surgical planning, diagnosis, and analysis, with magnetic resonance imaging (MRI) being the preferred modality due to its excellent visualisation of brain tissues. However, the wide intensity range of voxel values in MR scans often results in significant overlap between the density distributions of different tumour tissues, leading to reduced contrast and segmentation accuracy. This paper introduces a novel framework based on conditional generative adversarial networks (cGANs) aimed at enhancing the contrast of tumour subregions for both voxel-wise and region-wise segmentation approaches. We present two models: Enhancement and Segmentation GAN (ESGAN), which combines classifier loss with adversarial loss to predict central labels of input patches, and Enhancement GAN (EnhGAN), which generates high-contrast synthetic images with reduced inter-class overlap. These synthetic images are then fused with corresponding modalities to emphasise meaningful tissues while suppressing weaker ones. We also introduce a novel generator that adaptively calibrates voxel values within input patches, leveraging fully convolutional networks. Both models employ a multi-scale Markovian network as a GAN discriminator to capture local patch statistics and estimate the distribution of MR images in complex contexts. Experimental results on publicly available MR brain tumour datasets demonstrate the competitive accuracy of our models compared to current brain tumour segmentation techniques.

ProGleason-GAN: Conditional progressive growing GAN for prostatic cancer Gleason grade patch synthesis

BACKGROUND AND OBJECTIVE

Prostate cancer is one of the most common diseases affecting men. The main diagnostic and prognostic reference tool is the Gleason scoring system. An expert pathologist assigns a Gleason grade to a sample of prostate tissue. As this process is very time-consuming, some artificial intelligence applications were developed to automatize it. The training process is often confronted with insufficient and unbalanced databases which affect the generalisability of the models. Therefore, the aim of this work is to develop a generative deep learning model capable of synthesising patches of any selected Gleason grade to perform data augmentation on unbalanced data and test the improvement of classification models.

METHODOLOGY

The methodology proposed in this work consists of a conditional Progressive Growing GAN (ProGleason-GAN) capable of synthesising prostate histopathological tissue patches by selecting the desired Gleason Grade cancer pattern in the synthetic sample. The conditional Gleason Grade information is introduced into the model through the embedding layers, so there is no need to add a term to the Wasserstein loss function. We used minibatch standard deviation and pixel normalisation to improve the performance and stability of the training process.

RESULTS

The reality assessment of the synthetic samples was performed with the Frechet Inception Distance (FID). We obtained an FID metric of 88.85 for non-cancerous patterns, 81.86 for GG3, 49.32 for GG4 and 108.69 for GG5 after post-processing stain normalisation. In addition, a group of expert pathologists was selected to perform an external validation of the proposed framework. Finally, the application of our proposed framework improved the classification results in SICAPv2 dataset, proving its effectiveness as a data augmentation method.

CONCLUSIONS

ProGleason-GAN approach combined with a stain normalisation post-processing provides state-of-the-art results regarding Frechet's Inception Distance. This model can synthesise samples of non-cancerous patterns, GG3, GG4 or GG5. The inclusion of conditional information about the Gleason grade during the training process allows the model to select the cancerous pattern in a synthetic sample. The proposed framework can be used as a data augmentation method.

Enhanced dataset synthesis using conditional generative adversarial networks

Biomedical data acquisition, and reaching sufficient samples of participants are difficult and time ans effort consuming processes. On the other hand, the success rates of computer aided diagnosis (CAD) algorithms are sample and feature space depended. In this paper, conditional generative adversarial network (CGAN) based enhanced feature generation is proposed to synthesize large sample datasets having higher class separability. Twenty five percent of five medical datasets are used to train CGAN, and the synthetic datasets with any sample size are evaluated and compared to originals. Thus, new datasets can be generated with the help of the CGAN model and lower sample collection. It helps physicians decreasing sample collection processes, and it increases accuracy rates of the CAD systems using generated enhanced data with enhanced feature vectors. The synthesized datasets are classified using nearest neighbor, radial basis function support vector machine and artificial neural network to analyze the effectiveness of the proposed CGAN model.

LAC-GAN: Lesion attention conditional GAN for Ultra-widefield image synthesis

Automatic detection of retinal diseases based on deep learning technology and Ultra-widefield (UWF) images plays an important role in clinical practices in recent years. However, due to small lesions and limited data samples, it is not easy to train a detection-accurate model with strong generalization ability. In this paper, we propose a lesion attention conditional generative adversarial network (LAC-GAN) to synthesize retinal images with realistic lesion details to improve the training of the disease detection model. Specifically, the generator takes the vessel mask and class label as the conditional inputs, and processes the random Gaussian noise by a series of residual block to generate the synthetic images. To focus on pathological information, we propose a lesion feature attention mechanism based on random forest (RF) method, which constructs its reverse activation network to activate the lesion features. For discriminator, a weight-sharing multi-discriminator is designed to improve the performance of model by affine transformations. Experimental results on multi-center UWF image datasets demonstrate that the proposed method can generate retinal images with reasonable details, which helps to enhance the performance of the disease detection model.

Multi-Channel Generative Framework and Supervised Learning for Anomaly Detection in Surveillance Videos

Recently, most state-of-the-art anomaly detection methods are based on apparent motion and appearance reconstruction networks and use error estimation between generated and real information as detection features. These approaches achieve promising results by only using normal samples for training steps. In this paper, our contributions are two-fold. On the one hand, we propose a flexible multi-channel framework to generate multi-type frame-level features. On the other hand, we study how it is possible to improve the detection performance by supervised learning. The multi-channel framework is based on four Conditional GANs (CGANs) taking various type of appearance and motion information as input and producing prediction information as output. These CGANs provide a better feature space to represent the distinction between normal and abnormal events. Then, the difference between those generative and ground-truth information is encoded by Peak Signal-to-Noise Ratio (PSNR). We propose to classify those features in a classical supervised scenario by building a small training set with some abnormal samples of the original test set of the dataset. The binary Support Vector Machine (SVM) is applied for frame-level anomaly detection. Finally, we use Mask R-CNN as detector to perform object-centric anomaly localization. Our solution is largely evaluated on Avenue, Ped1, Ped2, and ShanghaiTech datasets. Our experiment results demonstrate that PSNR features combined with supervised SVM are better than error maps computed by previous methods. We achieve state-of-the-art performance for frame-level AUC on Ped1 and ShanghaiTech. Especially, for the most challenging Shanghaitech dataset, a supervised training model outperforms up to 9% the state-of-the-art an unsupervised strategy.