The rapid growth of social media has led to the widespread sharing of individual portrait images, which pose serious privacy risks due to the capabilities of automatic face recognition (AFR) systems for mass surveillance. Hence, protecting facial privacy against unauthorized AFR systems is essential. Inspired by the generation capability of the emerging diffusion models, recent methods employ diffusion models to generate adversarial face images for privacy protection. However, they suffer from the diffusion purification effect, leading to a low protection success rate (PSR). In this paper, we first propose learning unconditional embeddings to increase the learning capacity for adversarial modifications and then use them to guide the modification of the adversarial latent code to weaken the diffusion purification effect. Moreover, we integrate an identity-preserving structure to maintain structural consistency between the original and generated images, allowing human observers to recognize the generated image as having the same identity as the original. Extensive experiments conducted on two public datasets, i.e., CelebA-HQ and LADN, demonstrate the superiority of our approach. The protected faces generated by our method outperform those produced by existing facial privacy protection approaches in terms of transferability and natural appearance. The code is available at https://github.com/parham1998/FacialPrivacy-Protection

In many image-guided clinical approaches, tumor segmentation is a fundamental and critical step for locating tumor involvement. However, the scarcity of annotated data and the low contrast of medical imaging techniques make it challenging to accurately segment tumors from surrounding tissues using supervised learning methods. To address these issues, we propose a patch-based contrastive semi-supervised learning framework with wavelet guidance and adaptive data augmentation (PaCaS-WAA). Specifically, we apply patch-based contrast to maintain high-quality segmentation results with limited labels. Moreover, to exploit the discriminative information about subtle boundaries, we use the wavelet domain guides UNet for more edge details. Besides, to increase the diversity of unlabelled data, we propose an adaptive data augmentation strategy to augment the unlabelled data according to its Challenging Grade. Experimental results on two publicly available datasets of different modalities demonstrate that our method consistently outperform the state-of-the-art semi-supervised segmentation methods.

Nuclei segmentation is a critical step in the automated analysis of digitized microscopic images, which facilitates analysis of pathological images. The current state-of-the-art (SOTA) methods for nuclei segmentation require significant time and resources to provide pixel-level annotations for training. To reduce the labor-intensive annotation cost, we propose CCBox, a high-quality nuclei segmentation network that only requires bounding box annotations. We first generate hard pseudo labels for each nuclei within the bounding boxes using the traditional methods and then train a nuclei segmentation network with these hard pseudo labels. The major challenge lies in the presence of significant noise in the boundary regions of the nuclei due to the susceptibility of traditional methods to complex texture information in pathological images, impacting the model’s segmentation performance at the nucleus boundaries. To address this challenge, we propose the consistency constraint for similarity maps (CSM) strategy, which aggregates pixel features of the same semantics and enhances the discriminability of foreground and background features at the nucleus boundaries. Furthermore, to mitigate the overfitting of the model to noisy samples in the hard pseudo labels, we employ the hard-soft joint supervision strategy to supervise the student network. Extensive experiment results demonstrate that our CCBox significantly narrows the gap between box-supervised and fully-supervised nuclei segmentation methods.

We propose a novel three-stage Transformer-based methodology for entire cytopathological whole slide image (WSI) classification. The key idea is to leverage Transformer to extract the fine-grained lesion-level features and then progressively aggregate them into intermediate-grained patch-level features and coarse-grained WSI-level features for classification. Specifically, we first extract multi-scale lesion features from each patch image via Transformer-based lesion detection, and then adaptively aggregate the extracted lesion features into the corresponding patch feature with an MLP-Mixer. Finally, we select the most representative patch features and feed them into the Vision Transformer (ViT) for the final WSI-level classification. We collect a dataset consisting of 961 lung cytopathological WSIs of pleural effusions cytology specimens and conduct extensive experiments on it. The experimental results demonstrate that the proposed method outperforms existing state-of-the-art (SOTA) methods for cytopathological WSI classification.

Detecting tiny objects is a long-standing and critical problem in object detection, with broad real-world applications such as autonomous driving, surveillance, and medical diagnosis. Recent studies for tiny object detection often cause extra computational costs during inference due to introducing feature maps with increased resolution or additional network modules. This scarifies the inference speed for better detection accuracy and may heavily limit their availability to real-world applications. Therefore, this paper turns to knowledge distillation to improve the representation learning of a small model regarding both superior detection accuracy and fast inference speed. The masked scale-aware feature distillation and local attention distillation are proposed to address the critical issues in the distillation of tiny objects. Experimental results on two tiny benchmarks indicate that our method can bring noticeable performance gains to different detectors while keeping their original inference speeds. Our method also shows competitive performance compared to state-of-the-art methods for tiny object detection. Our code is available at https://github.com/haotianll/TinyKD.

Recently supervised contrastive learning (SCL) has achieved remarkable progress in semantic segmentation. Nevertheless, prior works have often necessitated a substantial number of samples to attain satisfactory performance, leading to a significant increase in training overhead. In this work, we leverage the idea of reweighting each pair to reduce the demand for large numbers of training samples in contrastive learning and propose a novel loss, dubbed confidence-aware contrastive (CAC) loss, which adaptively reweights each pair according to the predicted confidence for semantic segmentation. To alleviate the misalignment between supervised learning and contrastive learning, we further introduce an extra weight branch with a stop-gradient operator to generate the pair weights. Moreover, we present a confidence-aware marginal anchor sampling method for the calculation of supervised contrastive loss which focuses on marginal rather than the hardest pairs. Coupling with our method consistently improves the performance of various models (e.g. HRNet, OCRNet, SegFormer) on Cityscapes, ADE20K, PASCAL-Context, and COCO-Stuff datasets. Compared to existing SCL-based methods, the proposed method achieves competitive or even better results without relying on a memory bank or a large number of samples. Our code is at https://github.com/CVIU-CSU/Confidence-Aware-Contrastive-Loss.

Cervical abnormal cell detection is a challenging task as the morphological discrepancies between abnormal and normal cells are usually subtle. To determine whether a cervical cell is normal or abnormal, cytopathologists always take surrounding cells as references to identify its abnormality. To mimic these behaviors, we propose to explore contextual relationships to boost the performance of cervical abnormal cell detection. Specifically, both contextual relationships between cells and cell-to-global images are exploited to enhance features of each region of interest (RoI) proposal. Accordingly, two modules, dubbed as RoI-relationship attention module (RRAM) and global RoI attention module (GRAM), are developed and their combination strategies are also investigated. We establish a strong baseline by using Double-Head Faster R-CNN with a feature pyramid network (FPN) and integrate our RRAM and GRAM into it to validate the effectiveness of the proposed modules. Experiments conducted on a large cervical cell detection dataset reveal that the introduction of RRAM and GRAM both achieves better average precision (AP) than the baseline methods. Moreover, when cascading RRAM and GRAM, our method outperforms the state-of-the-art (SOTA) methods. Furthermore, we show that the proposed feature-enhancing scheme can facilitate image- and smear-level classification.

Fully automated cervical cancer grading on the level of Whole Slide Images (WSI) is a challenge task. As WSIs are in gigapixel resolution, it is impossible to train a deep classification neural network with the entire WSIs as inputs. To bypass this problem, we propose a two-stage learning framework. In detail, we propose to first learn patch-level deep pathological features for smear patches via a patch-level feature learning module, which is trained via leveraging the cell instance detection task. Then, we propose to learn WSI-level pathological features from patch-level features for cervical cancer grading. We conduct extensive experiments on our private dataset and make comparisons with rule-based cervical cancer grading methods. Experimental results demonstrate that our proposed deep feature-based WSI-level cervical cancer grading method achieves state-of-the-art performance.

As an important structure connecting the spine and lower limbs, the abnormal pelvis is one of the threats to human health worldwide, leading to millions of deaths every year. Although early diagnosis and treatment can greatly improve the chances of survival, it remains a major challenge, especially for pelvis fractures. Computer-aided pelvis segmentation (CPS) is a promising choice for the abnormal pelvis due to the great success of deep learning. However, when it comes to the abnormal pelvis diagnosis, the lack of training data has hampered the progress of CPS. To solve this problem, we have published a large-scale pelvis dataset, namely the Pelvis Computed tomography (CT) image (PCT14K) dataset. This dataset contains 14487 CT slices with the corresponding label for pelvis areas, while the existing largest public pelvis dataset part comes from existing data sets related to other organs with a lot of redundant information, and another part comes from the orthopedic hospital without a corresponding label. The proposed dataset enables the training of sophisticated segmentation networks for high-quality CPS. Some mainstream segmentation algorithms are trained and evaluated on the proposed PCT14K dataset and served as the baselines for future research. The published dataset will be available at https://github.com/YUAN-SIMING/PCT14K.

Unsupervised visual representation learning aims to learn general features from unlabelled data. Early methods design intra-image pretext tasks as learning targets and can be achieved with low computational overhead but unsatisfactory performance. Recent methods introduce contrastive learning and achieve surprising performance, but multiple views of training data are required in one batch, resulting in high computational overhead. To achieve competitive results to contrastive learning with low computational overhead, we propose a new unsupervised representation learning method with jigsaw clustering and classification as pretext tasks motivate the network to learn discriminative feature. To increase the data diversity, we propose to partition each training image into patches with random overlap, then randomly permute and stitch them into new training batch. Comparing with SOTAs, our method achieves state-of-the-art performance on both image classification/semi-classification on ImageNet and object detection on COCO.

Automated segmentation of hard exudates in colour fundus images is a challenge task due to issues of extreme class imbalance and enormous size variation. This paper aims to tackle these issues and proposes a dual-branch network with dual-sampling modulated Dice loss. It consists of two branches: large hard exudate biased segmentation branch and small hard exudate biased segmentation branch. Both of them are responsible for their own duties separately. Furthermore, we propose a dual-sampling modulated Dice loss for the training such that our proposed dual-branch network is able to segment hard exudates in different sizes. In detail, for the first branch, we use a uniform sampler to sample pixels from predicted segmentation mask for Dice loss calculation, which leads to this branch naturally be biased in favour of large hard exudates as Dice loss generates larger cost on misidentification of large hard exudates than small hard exudates. For the second branch, we use a re-balanced sampler to oversample hard exudate pixels and undersample background pixels for loss calculation. In this way, cost on misidentification of small hard exudates is enlarged, which enforces the parameters in the second branch fit small hard exudates well. Considering that large hard exudates are much easier to be correctly identified than small hard exudates, we propose an easy-to-difficult learning strategy by adaptively modulating the losses of two branches. We evaluate our proposed method on two public datasets and the results demonstrate that ours achieves state-of-the-art performance.

Abstract Retinal vessel segmentation is fundamental for the automatic retinal image analysis and ocular disease screening. This paper aims to learn a ground truth free feature aggregation strategy for the vessel segmentation. Five vesselness maps modelling the vessels'profile, appearance, and shape are first generated. Together, the histogram of the local binary pattern and the green colour are extracted. In each vesselness map, the pixels with large vesselness values are regarded as simple positive samples. The pixels with small vesselness values are regarded as simple negative samples, and the pixels with mediocre values are treated as difficult pixels. The simple positive samples and simple negative samples near the difficult pixels consist of the training dataset while the rest vesselness maps together with the local binary pattern histogram, and green colour channel are used as the features to learn a strong classifier. Then, without leveraging any ground truth, multiple kernel boosting is used to combine four support vector machine kernels to learn a strong vessel model for each image. Applying this learnt model to the pixels with mediocre values in the single vesselness map, their label will be determined. Totally, five strong vessel models are learnt. Finally, pixels with the majority supports from the strong vessel models are labelled as vessel pixels. The proposed method achieves accuracy of 94.83%, sensitivity of 72.59%, and specificity of 98.11% on DRIVE dataset, and accuracy of 95.51%, sensitivity of 78.09%, and specificity of 97.56% on STARE. It outperforms the state‐of‐the‐art unsupervised methods and achieves comparable performances to the supervised methods.

Segmentation of optic disc (OD) and optic cup (OC) is critical in automated fundus image analysis system. Existing state-of-the-arts focus on designing deep neural networks with one or multiple dense prediction branches. Such kind of designs ignore connections among prediction branches and their learning capacity is limited. To build connections among prediction branches, this paper introduces gradient boosting framework to deep classification model and proposes a gradient boosting network called BoostNet. Specifically, deformable side-output unit and aggregation unit with deep supervisions are proposed to learn base functions and expansion coefficients in gradient boosting framework. By stacking aggregation units in a deep-to-shallow manner, models’ performances are gradually boosted along deep to shallow stages. BoostNet achieves superior results to existing deep OD and OC segmentation networks on the public dataset ORIGA.

The microscopic urine sediment examination is a crucial part in the evaluation of renal and urinary tract diseases. Recently, there are emerging CNNs-based detectors to detect the urine sediment particles in an end-to-end manner. However, it is not very compatible to transfer CNNs-based detector directly from natural images application to microscopic images, especially in which small objects are in majority. This paper proposes a bidirectional context propagation network called BCPNet for urine sediment particle detection. In BCPNet, spatial details encoded by shallow convolutional layers are propagated upward to improve the localisation ability of deep features. On the contrary, high semantic information encoded by deep convolutional layers is propagated downward to enhance the distinctiveness of shallow features. With the refinement by convolutional block attention modules, the enriched features are more powerful to both localisation and classification. Experimental results on urine sediment particle dataset USE demonstrate effectiveness of the proposed BCPNet.

Color variations of histological images due to multi-factor hinder the performance of computer-aided diagnosis (CAD) systems. Previous stain normalization methods have achieved excellent results. While in practice, a multidomain stain normalization method is still be needed when more than two color variations exist in dataset. In this paper, we propose a multidomain stain normalization model inspired by MUNIT [1], with the idea of disentangling the representations of content and style. We assume that the latent space of histological images can be decomposed into domain-shared content space and domain-specific style space. The stain normalization aims to transfer the styles cross domains and maintain the contents. In addition, we propose to use the earth mover’s distance(EMD) to evaluate the effectiveness of stain normalization. We evaluate our approach against the state-of-the-art methods quantitatively and qualitatively.

Changes of retinal blood vessel are precursors of many serious diseases such as diabetic retinopathy, hypertension and cardiovascular diseases. Automatic segmentation of retinal blood vessels in the fundus image can better assist in the diagnosis of these diseases and has been studied by many researchers. However, the segmentation of pale vessel pixels remains a problem because of their low contrasts with surrounding pixels. This study proposes an improved multi‐scale line detector to segment retinal vessels. It computes the line responses of vessels in multi‐scale windows and takes the maximum as the response value, which can enhance the responses of pale vessel pixels near strong vessels or dark background pixels. Experimental results on the publicly available database DRIVE demonstrate that the proposed method can detect pale vessel pixels better. It achieves 75.28% in sensitivity and 94.47% in accuracy, which outperforms the state‐of‐the‐art unsupervised methods. Compared with the supervised methods it also gets better sensitivity and comparable accuracy.

Most existing salient object detection methods compute the saliency for pixels, patches, or superpixels by contrast. Such fine-grained contrast-based salient object detection methods are stuck with saliency attenuation of the salient object and saliency overestimation of the background when the image is complicated. To better compute the saliency for complicated images, we propose a hierarchical contour closure-based holistic salient object detection method, in which two saliency cues, i.e., closure completeness and closure reliability, are thoroughly exploited. The former pops out the holistic homogeneous regions bounded by completely closed outer contours, and the latter highlights the holistic homogeneous regions bounded by averagely highly reliable outer contours. Accordingly, we propose two computational schemes to compute the corresponding saliency maps in a hierarchical segmentation space. Finally, we propose a framework to combine the two saliency maps, obtaining the final saliency map. Experimental results on three publicly available datasets show that even each single saliency map is able to reach the state-of-the-art performance. Furthermore, our framework, which combines two saliency maps, outperforms the state of the arts. Additionally, we show that the proposed framework can be easily used to extend existing methods and further improve their performances substantially.