What Does a Model Really Look at?: Extracting Model-Oriented Concepts for Explaining Deep Neural Networks.

Model explainability is one of the crucial ingredients for building trustable AI systems, especially in the applications requiring reliability such as automated driving and diagnosis. Many explainability methods have been studied in the literature. Among many others, this article focuses on a research line that tries to visually explain a pre-trained image classification model such as Convolutional Neural Network by discovering concepts learned by the model, which is so-called the concept-based explanation. Previous concept-based explanation methods rely on the human definition of concepts (e.g., the Broden dataset) or semantic segmentation techniques like Slic (Simple Linear Iterative Clustering). However, we argue that the concepts identified by those methods may show image parts which are more in line with a human perspective or cropped by a segmentation method, rather than purely reflect a model's own perspective. We propose Model-Oriented Concept Extraction (MOCE), a novel approach to extracting key concepts based solely on a model itself, thereby being able to capture its unique perspectives which are not affected by any external factors. Experimental results on various pre-trained models confirmed the advantages of extracting concepts by truly representing the model's point of view.

An Interpretable Framework for Identifying Cerebral Microbleeds and Alzheimer's Disease Severity using Multimodal Data.

Cerebral microbleeds (CMBs) are tiny chronic brain haemorrhages that have been recognised as prognostic indicators for a number of acute cerebrovascular disorders, such as stroke, traumatic disorder, and Alzheimer's disease. For early-stage chronic disease diagnosis, it is challenging to automate the detection of CMBs and increase the reliability of prediction outputs. This study developed a system for identifying microbleeds in MRI images and gene expression data and determining the severity of Alzheimer's disease (AD). Initially, a spike neural network (SNN) and decision tree were utilised to identify microbleeds in AD from MRI images and gene expression respectively. However, the conclusions of these two methods cannot be interpreted due to the complexity of their internal processing steps. This study proposed two explainable artificial intelligence (XAI) methods for interpreting prediction outputs in an effort to boost reliability. Pixel density analysis (PDA) and probabilistic graphical model (PGM) explain the decision-making processes for MRI images and gene expression data for the diagnosis of microbleeds and the severity analysis of AD.

Mask-Transformer-Based Networks for Teeth Segmentation in Panoramic Radiographs.

Teeth segmentation plays a pivotal role in dentistry by facilitating accurate diagnoses and aiding the development of effective treatment plans. While traditional methods have primarily focused on teeth segmentation, they often fail to consider the broader oral tissue context. This paper proposes a panoptic-segmentation-based method that combines the results of instance segmentation with semantic segmentation of the background. Particularly, we introduce a novel architecture for instance teeth segmentation that leverages a dual-path transformer-based network, integrated with a panoptic quality (PQ) loss function. The model directly predicts masks and their corresponding classes, with the PQ loss function streamlining the training process. Our proposed architecture features a dual-path transformer block that facilitates bi-directional communication between the pixel path CNN and the memory path. It also contains a stacked decoder block that aggregates multi-scale features across different decoding resolutions. The transformer block integrates pixel-to-memory feedback attention, pixel-to-pixel self-attention, and memory-to-pixel and memory-to-memory self-attention mechanisms. The output heads process features to predict mask classes, while the final mask is obtained by multiplying memory path and pixel path features. When applied to the UFBA-UESC Dental Image dataset, our model exhibits a substantial improvement in segmentation performance, surpassing existing state-of-the-art techniques in terms of performance and robustness. Our research signifies an essential step forward in teeth segmentation and contributes to a deeper understanding of oral structures.

Efficient mutation screening for cervical cancers from circulating tumor DNA in blood.

BACKGROUND: Early diagnosis and continuous monitoring are necessary for an efficient management of cervical cancers (CC). Liquid biopsy, such as detecting circulating tumor DNA (ctDNA) from blood, is a simple, non-invasive method for testing and monitoring cancer markers. However, tumor-specific alterations in ctDNA have not been extensively investigated or compared to other circulating biomarkers in the diagnosis and monitoring of the CC. Therfore, Next-generation sequencing (NGS) analysis with blood samples can be a new approach for highly accurate diagnosis and monitoring of the CC. METHOD: Using a bioinformatics approach, we designed a panel of 24 genes associated with CC to detect and characterize patterns of somatic single-nucleotide variations, indels, and copy number variations. Our NGS CC panel covers most of the genes in The Cancer Genome Atlas (TCGA) as well as additional cancer driver and tumor suppressor genes. We profiled the variants in ctDNA from 24 CC patients who were being treated with systemic chemotherapy and local radiotherapy at the Jeonbuk National University Hospital, Korea. RESULT: Eighteen out of 24 genes in our NGS CC panel had mutations across the 24 CC patients, including somatic alterations of mutated genes (ZFHX3-83%, KMT2C-79%, KMT2D-79%, NSD1-67%, ATM-38% and RNF213-27%). We demonstrated that the RNF213 mutation could be used potentially used as a monitoring marker for response to chemo- and radiotherapy. CONCLUSION: We developed our NGS CC panel and demostrated that our NGS panel can be useful for the diagnosis and monitoring of the CC, since the panel detected the common somatic variations in CC patients and we observed how these genetic variations change according to the treatment pattern of the patient.

Combinatory Analysis of Cell-free and Circulating Tumor Cell DNAs Provides More Variants for Cancer Treatment.

BACKGROUND/AIM: Non-invasive biomarker detection using DNA from cell-free circulating DNA (cfDNA) and circulating tumor cells (ctcDNA) are emerging as they can be used for early diagnosis, prognosis and therapeutic target selection for cancer. However, cfDNA and ctcDNA from the same patient have not yet been compared extensively on how different the genetic characteristics of the two are in terms of the overlap between them. MATERIALS AND METHODS: The performance of a customized NGS panel was used to compare the variants found in the 20 pairs of cfDNA and ctcDNA from gynecological cancer patients. RESULTS: A genetic variant analysis revealed that there were only nine common overlapping variants out of 63 between the cfDNA and ctcDNA pairs, while 31 and 22 were unique to cfDNA and ctcDNA, respectively. CONCLUSION: A combinatory analysis of both cfDNA and CTCs from cancer patients can improve the sensitivity of liquid biopsies. These results are expected to provide better genetic target information for guiding clinical strategies for cancer.

MiR-195 and miR-497 suppress tumorigenesis in lung cancer by inhibiting SMURF2-induced TGF-ß receptor I ubiquitination.

SMURF2 is a member of the HECT family of E3 ubiquitin ligases that have important roles as a negative regulator of transforming growth factor-ß (TGF-ß) signaling through ubiquitin-mediated degradation of TGF-ß receptor I. However, the regulatory mechanism of SMURF2 is largely unknown. In this study, we identified that micro(mi)R-195 and miR-497 putatively target SMURF2 using several target prediction databases. Both miR-195 and miR-497 bind to the 3'-UTR of the SMURF2 mRNA and inhibit SMURF2 expression. Furthermore, miR-195 and miR-497 regulate SMURF2-dependent TßRI ubiquitination and cause the activation of the TGF-ß signaling pathway in lung cancer cells. Upregulation of miR-195 and miR-497 significantly reduced cell viability and colony formation through the activation of TGF-ß signaling. Interestingly, miR-195 and miR-497 also reduced the invasion ability of lung cancer cells when cells were treated with TGF-ß1. Subsequent in vivo studies in xenograft nude mice model revealed that miR-195 and miR-497 repress tumor growth. These findings demonstrate that miR-195 and miR-497 act as a tumor suppressor by suppressing ubiquitination-mediated degradation of TGF-ß receptors through SMURF2, and suggest that miR-195 and miR-497 are potential therapeutic targets for lung cancer.

An improvement of miRNA extraction efficiency in human plasma.

MicroRNAs (miRNAs) are short noncoding RNA molecules that control the expression of mRNAs associated with various biological processes. Therefore, deregulated miRNAs play important roles in the pathogenesis of diseases. Numerous studies are aimed at discovering biomarkers of diseases or determining miRNA functions by monitoring circulating miRNAs in various biological sources such as plasma and urine. However, the analysis of miRNA in such fluids presents problems related to accuracy and reproducibility because of their low levels in biological fluids. Therefore, better extraction kits and more sensitive detection systems have been developed for improved and reproducible analysis of circulating miRNAs. However, new extraction methods are also needed to improve the yield of miRNAs for their reliable analysis from biological fluids. The combination of yeast transfer RNA (tRNA) and glycogen as carrier molecules and incubation durations were optimized to maximize extraction efficiency. The extraction recovery using a combination of yeast tRNA and glycogen was approximately threefold more than that by using glycogen or yeast tRNA alone. In addition, reproducible and accurate analysis of miRNAs can be carried out after extraction using a combination of yeast tRNA and glycogen without an impact on plasma components. Graphical abstract Steps of miRNA extraction in plasma.

MIR-27a regulates the TGF-ß signaling pathway by targeting SMAD2 and SMAD4 in lung cancer.

The transforming growth factor-ß (TGF-ß) signaling pathway is associated with carcinogenesis and various biological processes. SMAD2 and SMAD4, which are putative tumor suppressors, have an important role in TGF-ß signaling. The aberrant expression of these genes is implicated in some cancers. However, the mechanisms of SMAD2 and SMAD4 dysregulation are poorly understood. In this study, we observed that miR-27a was upregulated in lung cancer cell lines and patients. In addition, SMAD2 and SMAD4 genes were identified as targets of miR-27a by several target prediction databases and experimental validation. Functional studies revealed that miR-27a overexpression decreased SMAD2 and SMAD4 mRNA and protein levels. Furthermore, miR-27a contributed to cell proliferation and invasion by inhibiting TGF-ß-induced cell cycle arrest. These results suggest that miR-27a may function as an oncogene by regulating SMAD2 and SMAD4 in lung cancer. Thus, miR-27a may be a potential target for cancer therapy.

Evaluation of inhibition of miRNA expression induced by anti-miRNA oligonucleotides.

MicroRNAs (miRNAs) are short RNA molecules that control the expression of mRNAs associated with various biological processes. Therefore, deregulated miRNAs play an important role in the pathogenesis of diseases. Numerous studies aimed at developing novel miRNA-based drugs or determining miRNA functions have been conducted by inhibiting miRNAs using anti-miRNA oligonucleotides (AMOs), which inhibit the function by hybridizing with miRNA. To increase the binding affinity and specificity to target miRNA, AMOs with various chemical modifications have been developed. Evaluating the potency of these various types of AMOs is an essential step in their development. In this study, we developed a capillary electrophoresis with laser-induced fluorescence (CE-LIF) method to evaluate the potency of AMOs by measuring changes in miRNA levels with fluorescence-labeled ssDNA probes using AMO-miR-23a, which inhibits miR-23a related to lung cancer. In order to eliminate interference by excess AMOs during hybridization of the ssDNA probe with the miR-23a, the concentration of the ssDNA probe was optimized. This newly developed method was used to compare the potency of two different modified AMOs. The data were supported by the results of a luciferase assay. This study demonstrated that CE-LIF analysis could be used to accurately evaluate AMO potency in biological samples.

Simultaneous detection of multiple microRNAs for expression profiles of microRNAs in lung cancer cell lines by capillary electrophoresis with dual laser-induced fluorescence.

MicroRNAs (miRNAs) are small, endogenous, single-stranded, noncoding RNAs. Circulating miRNAs are being considered as promising disease biomarkers. Indeed, single miRNAs have been associated with a wide variety of disease conditions and can target multiple mRNAs; therefore, several miRNAs may be simultaneously involved in disease progression and development. In this study, we developed a capillary electrophoresis with dual laser-induced fluorescence (CE with dual LIF) method using two color laser excitations for simultaneous determination of multiple miRNAs. Target miRNAs were hybridized with 6-FAM- or Cy5-labeled DNA probes for simultaneous determination of multiple miRNAs at excitation wavelengths of 488 and 635 nm. The hybridized miRNAs were separated using CE with dual LIF and detected within 13 min at excitation wavelengths of 488 and 635 nm without any interference or crosstalk. Additionally, the proposed approach was used successfully to detect and evaluate levels of several endogenous miRNAs from lung cancer cell lines. These results showed the potential of CE with dual LIF for fast, specific, simultaneous analysis of multiple miRNAs in cell extracts, biofluids, and tissues.

Enhanced extraction efficiency of miRNA from cells by addition of Triton X-100.

The determination of microRNA (miRNA) levels in biomaterials has become important for understanding their biological functions and for the diagnosis of various diseases. An effective extraction method is needed for maximizing the recovery of miRNAs from cells, while minimizing RNA degradation during the extraction because miRNAs present only approximately 0.01 % of total RNA. In this study, we used Triton X-100 (TX-100) to improve the extraction efficiency of miRNAs with TRIzol® reagent, which is a commonly used commercial microRNA isolation kit. The concentration of TX-100 and the incubation time after the addition of TX-100 were optimized to maximize the extraction efficiency. The extraction recovery by a combination of TX-100 and TRIzol® reagent was approximately 1.9-fold greater than that by the TRIzol® reagent alone. We have established a very effective extraction method for the extraction of low-abundance miRNAs in biological samples for the determination of miRNA levels in biomaterials.

Determination of micro-RNA in cardiomyoblast cells using CE with LIF detection.

Micro-RNAs (miRNAs) are small, endogenous, singlestranded, and noncoding RNAs. The miRNAs have been found to perform important functions in many cellular processes, such as development, proliferation, differentiation, and apoptosis. Circulating miRNAs have been proposed as emerging biomarkers in diseases such as cancer, diabetes, and cardiovascular disease including acute myocardial infarction (AMI). In this study, we developed CE with LIF (CE-LIF) using fluorescence-labeled DNA probe for determination of low abundance miRNA in cell extracts. The target miRNA is miRNA-499, a biomarker candidate of AMI with low abundance in biological samples. In order to measure the trace level of miRNA, we optimized the hybridization conditions such as hybridization time, temperature, and buffer solution. The highest fluorescence intensity of the hybridized miRNA-499 was found when hybridization was conducted at 40°C in 50 mM Tris-acetate (pH 8.0) buffer containing 50 mM NaCl, and 10 mM EDTA for 15 min. The hybridized miRNA-499 was detected in cultured H9c2 cardiomyoblast cells and the analysis of miRNA-499 was completed within 1 h using CE-LIF. These results showed the potential of CE for fast, specific, and sensitive high-throughput analysis of low-abundance miRNAs in cell extracts, biofluids, and tissues.