Jianhua Ma

Jae-Won Chung, Jeff J. Ma, Ruofan Wu et al.

As the adoption of Generative AI in real-world services grow explosively, energy has emerged as a critical bottleneck resource. However, energy remains a metric that is often overlooked, under-explored, or poorly understood in the context of building ML systems. We present the ML$.$ENERGY Benchmark, a benchmark suite and tool for measuring inference energy consumption under realistic service environments, and the corresponding ML$.$ENERGY Leaderboard, which have served as a valuable resource for those hoping to understand and optimize the energy consumption of their generative AI services. In this paper, we explain four key design principles for benchmarking ML energy we have acquired over time, and then describe how they are implemented in the ML$.$ENERGY Benchmark. We then highlight results from the early 2025 iteration of the benchmark, including energy measurements of 40 widely used model architectures across 6 different tasks, case studies of how ML design choices impact energy consumption, and how automated optimization recommendations can lead to significant (sometimes more than 40%) energy savings without changing what is being computed by the model. The ML$.$ENERGY Benchmark is open-source and can be easily extended to various customized models and application scenarios.

Peng Xu, Shengwu Xiong, Jiajun Zhang et al.

This paper reviews the MARS2 2025 Challenge on Multimodal Reasoning. We aim to bring together different approaches in multimodal machine learning and LLMs via a large benchmark. We hope it better allows researchers to follow the state-of-the-art in this very dynamic area. Meanwhile, a growing number of testbeds have boosted the evolution of general-purpose large language models. Thus, this year's MARS2 focuses on real-world and specialized scenarios to broaden the multimodal reasoning applications of MLLMs. Our organizing team released two tailored datasets Lens and AdsQA as test sets, which support general reasoning in 12 daily scenarios and domain-specific reasoning in advertisement videos, respectively. We evaluated 40+ baselines that include both generalist MLLMs and task-specific models, and opened up three competition tracks, i.e., Visual Grounding in Real-world Scenarios (VG-RS), Visual Question Answering with Spatial Awareness (VQA-SA), and Visual Reasoning in Creative Advertisement Videos (VR-Ads). Finally, 76 teams from the renowned academic and industrial institutions have registered and 40+ valid submissions (out of 1200+) have been included in our ranking lists. Our datasets, code sets (40+ baselines and 15+ participants' methods), and rankings are publicly available on the MARS2 workshop website and our GitHub organization page https://github.com/mars2workshop/, where our updates and announcements of upcoming events will be continuously provided.

Haotan Guo, Jianfei He, Jiayuan Ma et al.

Phonetic Cloaking Replacement (PCR), defined as the deliberate use of homophonic or near-homophonic variants to hide toxic intent, has become a major obstacle to Chinese content moderation. While this problem is well-recognized, existing evaluations predominantly rely on rule-based, synthetic perturbations that ignore the creativity of real users. We organize PCR into a four-way surface-form taxonomy and compile \ours, a dataset of 500 naturally occurring, phonetically cloaked offensive posts gathered from the RedNote platform. Benchmarking state-of-the-art LLMs on this dataset exposes a serious weakness: the best model reaches only an F1-score of 0.672, and zero-shot chain-of-thought prompting pushes performance even lower. Guided by error analysis, we revisit a Pinyin-based prompting strategy that earlier studies judged ineffective and show that it recovers much of the lost accuracy. This study offers the first comprehensive taxonomy of Chinese PCR, a realistic benchmark that reveals current detectors' limits, and a lightweight mitigation technique that advances research on robust toxicity detection.

Jeff J. Ma, Jae-Won Chung, Jisang Ahn et al.

We present Cornserve, an efficient online serving system for an emerging class of multimodal models called Any-to-Any models. Any-to-Any models accept combinations of text and multimodal data (e.g., image, video, audio) as input and also generate combinations of text and multimodal data as output, introducing request type, computation path, and computation scaling heterogeneity in model serving. Cornserve allows model developers to describe the computation graph of generic Any-to-Any models, which consists of heterogeneous components such as multimodal encoders, autoregressive models like Large Language Models (LLMs), and multimodal generators like Diffusion Transformers (DiTs). Given this, Cornserve's planner automatically finds an optimized deployment plan for the model, including whether and how to disaggregate the model into smaller components based on model and workload characteristics. Cornserve's distributed runtime then executes the model per the plan, efficiently handling Any-to-Any model heterogeneity during online serving. Evaluations show that Cornserve can efficiently serve diverse Any-to-Any models and workloads, delivering up to 3.81$\times$ throughput improvement and up to 5.79$\times$ tail latency reduction over existing solutions.

Runyu Lu, Shiqi He, Wenxuan Tan et al.

Diffusion Transformer (DiT) models excel at generating highquality images through iterative denoising steps, but serving them under strict Service Level Objectives (SLOs) is challenging due to their high computational cost, particularly at large resolutions. Existing serving systems use fixed degree sequence parallelism, which is inefficient for heterogeneous workloads with mixed resolutions and deadlines, leading to poor GPU utilization and low SLO attainment. In this paper, we propose step-level sequence parallelism to dynamically adjust the parallel degree of individual requests according to their deadlines. We present TetriServe, a DiT serving system that implements this strategy for highly efficient image generation. Specifically, TetriServe introduces a novel round-based scheduling mechanism that improves SLO attainment: (1) discretizing time into fixed rounds to make deadline-aware scheduling tractable, (2) adapting parallelism at the step level and minimize GPU hour consumption, and (3) jointly packing requests to minimize late completions. Extensive evaluation on state-of-the-art DiT models shows that TetriServe achieves up to 32% higher SLO attainment compared to existing solutions without degrading image quality.

Gaofeng Chen, Yaoduo Zhang, Li Huang et al.

Supervised deep-learning (SDL) techniques with paired training datasets have been widely studied for X-ray computed tomography (CT) image reconstruction. However, due to the difficulties of obtaining paired training datasets in clinical routine, the SDL methods are still away from common uses in clinical practices. In recent years, self-supervised deep-learning (SSDL) techniques have shown great potential for the studies of CT image reconstruction. In this work, we propose a self-supervised cross-task mutual learning (SS-CTML) framework for CT image reconstruction. Specifically, a sparse-view scanned and a limited-view scanned sinogram data are first extracted from a full-view scanned sinogram data, which results in three individual reconstruction tasks, i.e., the full-view CT (FVCT) reconstruction, the sparse-view CT (SVCT) reconstruction, and limited-view CT (LVCT) reconstruction. Then, three neural networks are constructed for the three reconstruction tasks. Considering that the ultimate goals of the three tasks are all to reconstruct high-quality CT images, we therefore construct a set of cross-task mutual learning objectives for the three tasks, in which way, the three neural networks can be self-supervised optimized by learning from each other. Clinical datasets are adopted to evaluate the effectiveness of the proposed framework. Experimental results demonstrate that the SS-CTML framework can obtain promising CT image reconstruction performance in terms of both quantitative and qualitative measurements.

Chenjie Zhou MD, Jianhua Ma Ph. D, Xiaoping Xu MD et al.

Objectives: Pancreatic cancer is a lethal disease, hard to diagnose and usually results in poor prognosis and high mortality. Developing an artificial intelligence (AI) algorithm to accurately and universally predict the early cancer risk of all kinds of pancreatic cancer is extremely important. We propose an ensemble AI algorithm to predict universally cancer risk of all kinds of pancreatic lesions with noncontrast CT. Methods: Our algorithm combines the radiomics method and a support tensor machine (STM) by the evidence reasoning (ER) technique to construct a binary classifier, called RadSTM-ER. RadSTM-ER takes advantage of the handcrafted features used in radiomics and learning features learned automatically by the STM from the CTs for presenting better characteristics of lesions. The patient cohort consisted of 135 patients with pathological diagnosis results where 97 patients had malignant lesions. Twenty-seven patients were randomly selected as independent test samples, and the remaining patients were used in a 5-fold cross validation experiment to confirm the hyperparameters, select optimal handcrafted features and train the model. Results: RadSTM-ER achieved independent test results: an area under the receiver operating characteristic curve of 0.8951, an accuracy of 85.19%, a sensitivity of 88.89%, a specificity of 77.78%, a positive predictive value of 88.89% and a negative predictive value of 77.78%. Conclusions: These results are better than the diagnostic performance of the five experimental radiologists, four conventional AI algorithms, which initially demonstrate the potential of noncontrast CT-based RadSTM-ER in cancer risk prediction for all kinds of pancreatic lesions.

Lisha Yao, Sui Li, Manman Zhu et al.

Energy-resolving computed tomography (ErCT) has the ability to acquire energy-dependent measurements simultaneously and quantitative material information with improved contrast-to-noise ratio. Meanwhile, ErCT imaging system is usually equipped with an advanced photon counting detector, which is expensive and technically complex. Therefore, clinical ErCT scanners are not yet commercially available, and they are in various stage of completion. This makes the researchers less accessible to the ErCT images. In this work, we investigate to produce ErCT images directly from existing energy-integrating CT (EiCT) images via deep neural network. Specifically, different from other networks that produce ErCT images at one specific energy, this model employs a unified generative adversarial network (uGAN) to concurrently train EiCT datasets and ErCT datasets with different energies and then performs image-to-image translation from existing EiCT images to multiple ErCT image outputs at various energy bins. In this study, the present uGAN generates ErCT images at 70keV, 90keV, 110keV, and 130keV simultaneously from EiCT images at140kVp. We evaluate the present uGAN model on a set of over 1380 CT image slices and show that the present uGAN model can produce promising ErCT estimation results compared with the ground truth qualitatively and quantitatively.

Shulong Li, Panpan Xu, Bin Li et al.

To predict lung nodule malignancy with a high sensitivity and specificity, we propose a fusion algorithm that combines handcrafted features (HF) into the features learned at the output layer of a 3D deep convolutional neural network (CNN). First, we extracted twenty-nine handcrafted features, including nine intensity features, eight geometric features, and twelve texture features based on grey-level co-occurrence matrix (GLCM) averaged from thirteen directions. We then trained 3D CNNs modified from three state-of-the-art 2D CNN architectures (AlexNet, VGG-16 Net and Multi-crop Net) to extract the CNN features learned at the output layer. For each 3D CNN, the CNN features combined with the 29 handcrafted features were used as the input for the support vector machine (SVM) coupled with the sequential forward feature selection (SFS) method to select the optimal feature subset and construct the classifiers. The fusion algorithm takes full advantage of the handcrafted features and the highest level CNN features learned at the output layer. It can overcome the disadvantage of the handcrafted features that may not fully reflect the unique characteristics of a particular lesion by combining the intrinsic CNN features. Meanwhile, it also alleviates the requirement of a large scale annotated dataset for the CNNs based on the complementary of handcrafted features. The patient cohort includes 431 malignant nodules and 795 benign nodules extracted from the LIDC/IDRI database. For each investigated CNN architecture, the proposed fusion algorithm achieved the highest AUC, accuracy, sensitivity, and specificity scores among all competitive classification models.

Ji He, Jianhua Ma

Radon transform is widely used in physical and life sciences and one of its major applications is the X-ray computed tomography (X-ray CT), which is significant in modern health examination. The Radon inversion or image reconstruction is challenging due to the potentially defective radon projections. Conventionally, the reconstruction process contains several ad hoc stages to approximate the corresponding Radon inversion. Each of the stages is highly dependent on the results of the previous stage. In this paper, we propose a novel unified framework for Radon inversion via deep learning (DL). The Radon inversion can be approximated by the proposed framework with an end-to-end fashion instead of processing step-by-step with multiple stages. For simplicity, the proposed framework is short as iRadonMap (inverse Radon transform approximation). Specifically, we implement the iRadonMap as an appropriative neural network, of which the architecture can be divided into two segments. In the first segment, a learnable fully-connected filtering layer is used to filter the radon projections along the view-angle direction, which is followed by a learnable sinusoidal back-projection layer to transfer the filtered radon projections into an image. The second segment is a common neural network architecture to further improve the reconstruction performance in the image domain. The iRadonMap is overall optimized by training a large number of generic images from ImageNet database. To evaluate the performance of the iRadonMap, clinical patient data is used. Qualitative results show promising reconstruction performance of the iRadonMap.

Hao Zhang, Jing Wang, Jianhua Ma et al.

Statistical image reconstruction (SIR) methods have shown potential to substantially improve the image quality of low-dose X-ray computed tomography (CT) as compared to the conventional filtered back-projection (FBP) method for various clinical tasks. According to the maximum a posterior (MAP) estimation, the SIR methods can be typically formulated by an objective function consisting of two terms: (1) data-fidelity (or equivalently, data-fitting or data-mismatch) term modeling the statistics of projection measurements, and (2) regularization (or equivalently, prior or penalty) term reflecting prior knowledge or expectation on the characteristics of the image to be reconstructed. Existing SIR methods for low-dose CT can be divided into two groups: (1) those that use calibrated transmitted photon counts (before log-transform) with penalized maximum likelihood (pML) criterion, and (2) those that use calibrated line-integrals (after log-transform) with penalized weighted least-squares (PWLS) criterion. Accurate statistical modeling of the projection measurements is a prerequisite for SIR, while the regularization term in the objective function also plays a critical role for successful image reconstruction. This paper reviews several statistical models on CT projection measurements and various regularization strategies incorporating prior knowledge or expected properties of the image to be reconstructed, which together formulate the objective function of the SIR methods for low-dose X-ray CT.