Towards Benchmarking AI Explainability

Brief hospital course (BHC) summaries are clinical documents that summarize a patient's hospital stay. While large language models (LLMs) depict remarkable capabilities in automating real-world tasks, their capabilities for healthcare applications such as synthesizing BHCs from clinical notes have not been shown. We introduce a novel preprocessed dataset, the MIMIC-IV-BHC, encapsulating clinical note and BHC pairs to adapt LLMs for BHC synthesis. Furthermore, we introduce a benchmark of the summarization performance of 2 general-purpose LLMs and 3 healthcare-adapted LLMs. Using clinical notes as input, we apply prompting-based (using in-context learning) and fine-tuning-based adaptation strategies to 3 open-source LLMs (Clinical-T5-Large, Llama2-13B, and FLAN-UL2) and 2 proprietary LLMs (Generative Pre-trained Transformer [GPT]-3.5 and GPT-4). We evaluate these LLMs across multiple context-length inputs using natural language similarity metrics. We further conduct a clinical study with 5 clinicians, comparing clinician-written and LLM-generated BHCs across 30 samples, focusing on their potential to enhance clinical decision-making through improved summary quality. We compare reader preferences for the original and LLM-generated summary using Wilcoxon signed-rank tests. We further request optional qualitative feedback from clinicians to gain deeper insights into their preferences, and we present the frequency of common themes arising from these comments. The Llama2-13B fine-tuned LLM outperforms other domain-adapted models given quantitative evaluation metrics of Bilingual Evaluation Understudy (BLEU) and Bidirectional Encoder Representations from Transformers (BERT)-Score. GPT-4 with in-context learning shows more robustness to increasing context lengths of clinical note inputs than fine-tuned Llama2-13B. Despite comparable quantitative metrics, the reader study depicts a significant preference for summaries generated by GPT-4 with in-context learning compared to both Llama2-13B fine-tuned summaries and the original summaries (P<.001), highlighting the need for qualitative clinical evaluation. We release a foundational clinically relevant dataset, the MIMIC-IV-BHC, and present an open-source benchmark of LLM performance in BHC synthesis from clinical notes. We observe high-quality summarization performance for both in-context proprietary and fine-tuned open-source LLMs using both quantitative metrics and a qualitative clinical reader study. Our research effectively integrates elements from the data assimilation pipeline: our methods use (1) clinical data sources to integrate, (2) data translation, and (3) knowledge creation, while our evaluation strategy paves the way for (4) deployment.

This paper presents a novel quantitative metric for comparison of 3D mesh segmentations, Ultimate Measurement Accuracy, basing on the screening data sets, to support our quantitative comparisons of seven recently published mesh segmentation algorithms; and the experiment results suggest that, our metrics is robust to degenerative segmentation and hierarchical refinement, stable to the imprecision of cut boundaries.

Simplification of feature-based 3D CAD assembly data of ship and offshore equipment using quantitative evaluation metrics

Deep learning (DL) has been applied in proofs of concept across biomedical imaging, including across modalities and medical specialties. Labeled data are critical to training and testing DL models, but human expert labelers are limited. In addition, DL traditionally requires copious training data, which is computationally expensive to process and iterate over. Consequently, it is useful to prioritize using those images that are most likely to improve a model's performance, a practice known as instance selection. The challenge is determining how best to prioritize. It is natural to prefer straightforward, robust, quantitative metrics as the basis for prioritization for instance selection. However, in current practice, such metrics are not tailored to, and almost never used for, image datasets. To address this problem, we introduce ENRICH-Eliminate Noise and Redundancy for Imaging Challenges-a customizable method that prioritizes images based on how much diversity each image adds to the training set. First, we show that medical datasets are special in that in general each image adds less diversity than in nonmedical datasets. Next, we demonstrate that ENRICH achieves nearly maximal performance on classification and segmentation tasks on several medical image datasets using only a fraction of the available images and without up-front data labeling. ENRICH outperforms random image selection, the negative control. Finally, we show that ENRICH can also be used to identify errors and outliers in imaging datasets. ENRICH is a simple, computationally efficient method for prioritizing images for expert labeling and use in DL.

This study presents a new approach based on artificial neural networks for generating one biometric feature (faces) from another (only fingerprints). An automatic and intelligent system was designed and developed to analyze the relationships among fingerprints and faces and also to model and to improve the existence of the relationships. The new proposed system is the first study that generates all parts of the face including eyebrows, eyes, nose, mouth, ears and face border from only fingerprints. It is also unique and different from similar studies recently presented in the literature with some superior features. The parameter settings of the system were achieved with the help of Taguchi experimental design technique. The performance and accuracy of the system have been evaluated with 10-fold cross validation technique using qualitative evaluation metrics in addition to the expanded quantitative evaluation metrics. Consequently, the results were presented on the basis of the combination of these objective and subjective metrics for illustrating the qualitative properties of the proposed methods as well as a quantitative evaluation of their performances. Experimental results have shown that one biometric feature can be determined from another. These results have once more indicated that there is a strong relationship between fingerprints and faces.

Segmentation of breast ultrasound images is a crucial and challenging task in computer-aided diagnosis systems. Accurately segmenting masses in benign and malignant cases and identifying regions with no mass is a primary objective in breast ultrasound image segmentation. Deep learning (DL) has emerged as a powerful tool in medical image segmentation, revolutionizing how medical professionals analyze and interpret complex imaging data. The UNet architecture is a highly regarded and widely used DL model in medical image segmentation. Its distinctive architectural design and exceptional performance have made it popular among researchers. With the increase in data and model complexity, optimization and fine-tuning models play a vital and more challenging role than before. This paper presents a comparative study evaluating the effect of image preprocessing and different optimization techniques and the importance of fine-tuning different UNet segmentation models for breast ultrasound images. Optimization and fine-tuning techniques have been applied to enhance the performance of UNet, Sharp UNet, and Attention UNet. Building upon this progress, we designed a novel approach by combining Sharp UNet and Attention UNet, known as Sharp Attention UNet. Our analysis yielded the following quantitative evaluation metrics for the Sharp Attention UNet: the Dice coefficient, specificity, sensitivity, and F1 score values obtained were 0.93, 0.99, 0.94, and 0.94, respectively. In addition, McNemar’s statistical test was applied to assess significant differences between the approaches. Across a number of measures, our proposed model outperformed all other models, resulting in improved breast lesion segmentation.

The integration of Explainable AI (XAI) into healthcare promises greater transparency and interpretability of machine learning models, enabling clinicians to understand predictions and make more reliable medical decisions. Yet, the robustness of XAI methods remains uncertain, as small input perturbations can drastically change their explanations, posing critical risks in clinical settings where they may lead to misdiagnoses or inappropriate treatment. Motivated by the central role of XAI in healthcare decision-making, this paper examines its robustness in the presence of data corruption. We systematically evaluate the stability of widely used XAI techniques against both naturally occurring noise (e.g., JPEG compression) and adversarial manipulations that alter explanations without affecting model predictions. To this end, we introduce a set of evaluation metrics that capture complementary aspects of explanation stability, ranging from pixel-level consistency to spatial coherence, and propose a protocol for assessing the resilience of XAI methods across diverse perturbation sources. Our analysis spans three medical imaging datasets, various convolutional and transformer models, and ten post-hoc XAI methods, including Grad-CAM++ for convolutional networks and LibraGrad for vision transformers. We find that current XAI techniques are often unstable, even under imperceptible perturbations. For adversarial noise, a clear set of robust methods emerges, whereas for natural noise, performance varies, with some methods maintaining spatial stability and others preserving pixel-wise consistency. All results together highlight the need for multi-perspective evaluation when selecting XAI techniques in practice.

The federated learning (FL) paradigm is well-suited for the field of medical image analysis, as it can effectively cope with machine learning on isolated multi-center data while protecting the privacy of participating parties. However, current research on optimization algorithms in FL often focuses on limited datasets and scenarios, primarily centered around natural images, with insufficient comparative experiments in medical contexts. In this work, we conduct a comprehensive evaluation of several state-of-the-art FL algorithms in the context of medical imaging. We conduct a fair comparison of classification models trained using various FL algorithms across multiple medical imaging datasets. Additionally, we evaluate system performance metrics, such as communication cost and computational efficiency, while considering different FL architectures. Our findings show that medical imaging datasets pose substantial challenges for current FL optimization algorithms. No single algorithm consistently delivers optimal performance across all medical FL scenarios, and many optimization algorithms may under-perform when applied to these datasets. Our experiments provide a benchmark and guidance for future research and application of FL in medical imaging contexts. Furthermore, we propose an efficient and robust method that combines generative techniques using denoising diffusion probabilistic models with label smoothing to augment datasets, widely enhancing the performance of FL on classification tasks across various medical imaging datasets.

Deep unfolding networks have obtained satisfactory performance in the pansharpening task owing to their sufficient interpretability. Inspired by the back-projection (BP) mechanism, we propose a BP-driven model, spatial-spectral dual back-project network (S <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> DBPN), to fuse the low spatial resolution multispectral (LR MS) and the high spatial resolution panchromatic (PAN) images by exploiting the BP in spatial and spectral domains. Specifically, the proposed S <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> DBPN is made up of a spatial BP network, a spectral BP network, and a reconstruction network. In the spatial BP network, spatial down- and up-projection modules are derived from BP, which is responsible for the projection of the LR MS image into the spatial domain. By analogy with the spatial BP, we reformulate the degradation between high spatial resolution multispectral (HR MS) and PAN images as spectral down- and up-projections. Then, the spectral BP network is constructed for the projection of the PAN image along the channel dimension. Finally, the features from spatial and spectral BP networks are integrated to produce the desired HR MS image through the reconstruction network. Compared to the state-of-the-art methods, extensive experiments on QuickBird, GeoEye-1, and WorldView-2 datasets demonstrate that our S <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> DBPN produces better HR MS images in terms of qualitative and quantitative evaluation metrics. The code of S <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> DBPN is released at: https://github.com/RSMagneto/S2DBPN.

The fusion of traditional HSI mainly involves the fusion of HSI with MSI or PAN. Nevertheless, the combination of the former is constrained by the inadequate spatial resolution of MSI, and therefore the enhancement of the spatial resolution of HSI is inadequate. The latter is constrained by the substantial disparity in spectral information between PAN and HSI, leading to pronounced distortion of high-resolution HSI. To address this issue, this paper proposes IFNPMH, which is an integrated fusion network for used for the fusion of HSI, MSI, and PAN. The IFNPMH algorithm is capable of efficiently extracting spatial information from MSI and PAN, as well as spectral information from HSI. It combines the spatial and spectral information to produce fused images that possess precise spectral information and distinct spatial details. Specifically, IFNPMH incorporates the CBAM attention mechanism module to enhance feature representation capabilities. Finally, this article proposes a loss function that combines spectral and spatial information loss in IFNPMH training. The effectiveness of the proposed IFNPMH is demonstrated by extensive testing in comparison to other sophisticated approaches. This model has shown promising performance in terms of qualitative visual effects and quantitative evaluation metrics.

Accurate and reliable short-term runoff prediction plays a pivotal role in water resource management, agriculture, and flood control, enabling decision-makers to implement timely and effective measures to enhance water use efficiency and minimize losses. To further enhance the accuracy of runoff prediction, this study proposes a FA-LSTM model that integrates the Firefly algorithm (FA) with the long short-term memory neural network (LSTM). The research focuses on historical daily runoff data from the Dahuangjiangkou and Wuzhou Hydrology Stations in the Xijiang River Basin. The FA-LSTM model is compared with RNN, LSTM, GRU, SVM, and RF models. The FA-LSTM model was used to carry out the generalization experiment in Qianjiang, Wuxuan, and Guigang hydrology stations. Additionally, the study analyzes the performance of the FA-LSTM model across different forecasting horizons (1–5 days). Four quantitative evaluation metrics—mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R2), and Kling–Gupta efficiency coefficient (KGE)—are utilized in the evaluation process. The results indicate that: (1) Compared to RNN, LSTM, GRU, SVM, and RF models, the FA-LSTM model exhibits the best prediction performance, with daily runoff prediction determination coefficients (R2) reaching as high as 0.966 and 0.971 at the Dahuangjiangkou and Wuzhou Stations, respectively, and the KGE is as high as 0.965 and 0.960, respectively. (2) FA-LSTM model was used to conduct generalization tests at Qianjiang, Wuxuan and Guigang hydrology stations, and its R2 and KGE are 0.96 or above, indicating that the model has good adaptability in different hydrology stations and strong robustness. (3) As the prediction period extends, the R2 and KGE of the FA-LSTM model show a decreasing trend, but the whole model still showed feasible forecasting ability. The FA-LSTM model introduced in this study presents an effective new approach for daily runoff prediction.

Accurate and high-resolution weather radar data reflecting detailed structure information of radar echo plays an important role in analysis and forecast of extreme weather. Typically, this is done using interpolation schemes, which only use several neighboring data values for computational approximation to get the estimated, resulting the loss of intense echo information. Focus on this limitation, a superresolution reconstruction algorithm of weather radar data based on adaptive sparse domain selection (ASDS) is proposed in this article. First, the ASDS algorithm gets a compact dictionary by learning the precollected data of model weather radar echo patches. Second, the most relevant subdictionaries are adaptively select for each low-resolution echo patches during the spare coding. Third, two adaptive regularization terms are introduced to further improve the reconstruction effect of the edge and intense echo information of the radar echo. Experimental results show that the ASDS algorithm substantially outperforms interpolation methods for ×2 and ×4 reconstruction in terms of both visual quality and quantitative evaluation metrics.

Segmentation of organs-at-risk in cervical cancer CT images with a convolutional neural network

This work introduces and evaluates a new multiple-component stochastic mortality model. Our proposal is based on a parameter estimation methodology that aims to reveal significant and distinct age clusters by identifying the optimal number of incorporated period and cohort effects. Our methodology adopts sparse principal component analysis and generalized linear models (GLMs), first introduced in Hatzopoulos and Haberman (2011), and it incorporates several novelties. Precisely, our approach is driven by the unexplained variance ratio (UVR) metric to maximize the captured variance of the mortality data and to regulate the sparsity of the model with the aim of acquiring distinct and significant stochastic components. In this way, our model gains a highly informative structure in an efficient way, and it is able to designate an identified mortality trend to a unique age cluster. We also provide an extensive experimental testbed to evaluate the efficiency of the proposed model in terms of fitting and forecasting performance over several datasets (Greece, England and Wales, France, and Japan), and we compare our results to those of well-known mortality models (Lee-Carter, Renshaw-Haberman, Currie, and Plat). Our model is able to achieve high scores over diverse qualitative and quantitative evaluation metrics and outperforms the rest of the models in the majority of the experiments. Our results show the beneficial characteristics of the proposed model and come into agreement with well-established findings in the mortality literature.

Employee performance evaluation in many organizations often faces challenges due to numerous assessment criteria and potential subjectivity in the decision-making process, making the evaluation results less consistent and objective. Multi-Criteria Decision Making (MCDM) methods have been widely used to address this problem; however, previous approaches generally still rely on subjective weight determination and do not fully consider the stability of results against data variation. Therefore, this study aims to develop a more objective and stable decision-making model by integrating the LODECI method to determine criteria weights based on data and the SPOTIS method to rank alternatives based on their distance from the ideal solution. Five evaluation criteria are used, namely productivity, work quality, discipline, teamwork, and responsibility, with data collected from eight employees as alternatives. The analysis process was carried out through the stages of constructing a decision matrix, calculating criterion weights using LODECI, and ranking using SPOTIS which produced a total distance value as a quantitative evaluation metric. The research results show that GS Employee achieved the smallest distance value of 0.058, thus ranking first, followed by CR Employee with a value of 0.086 and AN Employee with a value of 0.321. These findings indicate that the proposed model is capable of providing more measurable and consistent evaluation results. The main contribution of this study lies in the integration of objective weighting and ideal-solution-based ranking methods supported by sensitivity analysis, thereby producing a performance evaluation system that is more reliable, transparent, and robust compared to previous approaches.