Assessing generalisability of deep learning-based polyp detection and segmentation methods through a computer vision challenge.

Polyps are well-known cancer precursors identified by colonoscopy. However, variability in their size, appearance, and location makes the detection of polyps challenging. Moreover, colonoscopy surveillance and removal of polyps are highly operator-dependent procedures and occur in a highly complex organ topology. There exists a high missed detection rate and incomplete removal of colonic polyps. To assist in clinical procedures and reduce missed rates, automated methods for detecting and segmenting polyps using machine learning have been achieved in past years. However, the major drawback in most of these methods is their ability to generalise to out-of-sample unseen datasets from different centres, populations, modalities, and acquisition systems. To test this hypothesis rigorously, we, together with expert gastroenterologists, curated a multi-centre and multi-population dataset acquired from six different colonoscopy systems and challenged the computational expert teams to develop robust automated detection and segmentation methods in a crowd-sourcing Endoscopic computer vision challenge. This work put forward rigorous generalisability tests and assesses the usability of devised deep learning methods in dynamic and actual clinical colonoscopy procedures. We analyse the results of four top performing teams for the detection task and five top performing teams for the segmentation task. Our analyses demonstrate that the top-ranking teams concentrated mainly on accuracy over the real-time performance required for clinical applicability. We further dissect the devised methods and provide an experiment-based hypothesis that reveals the need for improved generalisability to tackle diversity present in multi-centre datasets and routine clinical procedures.

Medical federated learning with joint graph purification for noisy label learning.

In terms of increasing privacy issues, Federated Learning (FL) has received extensive attention in medical imaging. Through collaborative training, FL can produce superior diagnostic models with global knowledge, while preserving private data locally. In practice, medical diagnosis suffers from intra-/inter-observer variability, thus label noise is inevitable in dataset preparation. Different from existing studies on centralized datasets, the label noise problem in FL scenarios confronts more challenges, due to data inaccessibility and even noise heterogeneity. In this work, we propose a federated framework with joint Graph Purification (FedGP) to address the label noise in FL through server and clients collaboration. Specifically, to overcome the impact of label noise on local training, we first devise a noisy graph purification on the client side to generate reliable pseudo labels by progressively expanding the purified graph with topological knowledge. Then, we further propose a graph-guided negative ensemble loss to exploit the topology of the client-side purified graph with robust complementary supervision against label noise. Moreover, to address the FL label noise with data silos, we propose a global centroid aggregation on the server side to produce a robust classifier with global knowledge, which can be optimized collaboratively in the FL framework. Extensive experiments are conducted on endoscopic and pathological images with the comparison under the homogeneous, heterogeneous, and real-world label noise for medical FL. Among these diverse noisy FL settings, our FedGP framework significantly outperforms denoising and noisy FL state-of-the-arts by a large margin. The source code is available at https://github.com/CUHK-AIM-Group/FedGP.

Decoupled Unbiased Teacher for Source-Free Domain Adaptive Medical Object Detection.

Source-free domain adaptation (SFDA) aims to adapt a lightweight pretrained source model to unlabeled new domains without the original labeled source data. Due to the privacy of patients and storage consumption concerns, SFDA is a more practical setting for building a generalized model in medical object detection. Existing methods usually apply the vanilla pseudo-labeling technique, while neglecting the bias issues in SFDA, leading to limited adaptation performance. To this end, we systematically analyze the biases in SFDA medical object detection by constructing a structural causal model (SCM) and propose an unbiased SFDA framework dubbed decoupled unbiased teacher (DUT). Based on the SCM, we derive that the confounding effect causes biases in the SFDA medical object detection task at the sample level, feature level, and prediction level. To prevent the model from emphasizing easy object patterns in the biased dataset, a dual invariance assessment (DIA) strategy is devised to generate counterfactual synthetics. The synthetics are based on unbiased invariant samples in both discrimination and semantic perspectives. To alleviate overfitting to domain-specific features in SFDA, we design a cross-domain feature intervention (CFI) module to explicitly deconfound the domain-specific prior with feature intervention and obtain unbiased features. Besides, we establish a correspondence supervision prioritization (CSP) strategy for addressing the prediction bias caused by coarse pseudo-labels by sample prioritizing and robust box supervision. Through extensive experiments on multiple SFDA medical object detection scenarios, DUT yields superior performance over previous state-of-the-art unsupervised domain adaptation (UDA) and SFDA counterparts, demonstrating the significance of addressing the bias issues in this challenging task. The code is available at https://github.com/CUHK-AIM-Group/Decoupled-Unbiased-Teacher.

SIGMA++: Improved Semantic-Complete Graph Matching for Domain Adaptive Object Detection.

Domain Adaptive Object Detection (DAOD) generalizes the object detector from an annotated domain to a label-free novel one. Recent works estimate prototypes (class centers) and minimize the corresponding distances to adapt the cross-domain class conditional distribution. However, this prototype-based paradigm 1) fails to capture the class variance with agnostic structural dependencies, and 2) ignores the domain-mismatched classes with a sub-optimal adaptation. To address these two challenges, we propose an improved SemantIc-complete Graph MAtching framework, dubbed SIGMA++, for DAOD, completing mismatched semantics and reformulating adaptation with hypergraph matching. Specifically, we propose a Hypergraphical Semantic Completion (HSC) module to generate hallucination graph nodes in mismatched classes. HSC builds a cross-image hypergraph to model class conditional distribution with high-order dependencies and learns a graph-guided memory bank to generate missing semantics. After representing the source and target batch with hypergraphs, we reformulate domain adaptation with a hypergraph matching problem, i.e., discovering well-matched nodes with homogeneous semantics to reduce the domain gap, which is solved with a Bipartite Hypergraph Matching (BHM) module. Graph nodes are used to estimate semantic-aware affinity, while edges serve as high-order structural constraints in a structure-aware matching loss, achieving fine-grained adaptation with hypergraph matching. The applicability of various object detectors verifies the generalization of SIGMA++, and extensive experiments on nine benchmarks show its state-of-the-art performance on both AP 50 and adaptation gains.

GRAB-Net: Graph-Based Boundary-Aware Network for Medical Point Cloud Segmentation.

Point cloud segmentation is fundamental in many medical applications, such as aneurysm clipping and orthodontic planning. Recent methods mainly focus on designing powerful local feature extractors and generally overlook the segmentation around the boundaries between objects, which is extremely harmful to the clinical practice and degenerates the overall segmentation performance. To remedy this problem, we propose a GRAph-based Boundary-aware Network (GRAB-Net) with three paradigms, Graph-based Boundary-perception Module (GBM), Outer-boundary Context-assignment Module (OCM), and Inner-boundary Feature-rectification Module (IFM), for medical point cloud segmentation. Aiming to improve the segmentation performance around boundaries, GBM is designed to detect boundaries and interchange complementary information inside semantic and boundary features in the graph domain, where semantics-boundary correlations are modelled globally and informative clues are exchanged by graph reasoning. Furthermore, to reduce the context confusion that degenerates the segmentation performance outside the boundaries, OCM is proposed to construct the contextual graph, where dissimilar contexts are assigned to points of different categories guided by geometrical landmarks. In addition, we advance IFM to distinguish ambiguous features inside boundaries in a contrastive manner, where boundary-aware contrast strategies are proposed to facilitate the discriminative representation learning. Extensive experiments on two public datasets, IntrA and 3DTeethSeg, demonstrate the superiority of our method over state-of-the-art methods.

Spatiotemporal variation of climate of different flanks and elevations of the Qinling-Daba mountains in China during 1969-2018.

Climate change exhibits great variation on different flanks and at different elevations in the same mountain range. To investigate the complexity of the geographic patterns of climate and phenology in the Qinling-Daba mountains (QDM), in the North-South transition zone of China, this study analyzed the spatiotemporal distribution characteristics of daily air temperature and precipitation data measured at 118 national weather stations (1969-2018). The principal findings were as follows. (1) Overall, a significant trend of warming was detected in all seasons over the past 50 years, with rates of increase of 0.347, 0.125, 0.200 and 0.302 °C/10a, in spring, summer, autumn and winter, respectively. Precipitation did not show significant variation at most stations in different seasons. (2) The rising rate of air temperature varied considerably between different flanks. Generally, air temperature change on northern flanks was greater than on southern flanks in all seasons. The tendency of air temperature change was greater in spring and winter than in summer and autumn on different flanks in the QDM. (3) The rate of increase in high-elevation regions was greater than in low-elevation regions in summer, autumn and winter, e.g., 0.440, 0.390 and 0.456 °C/10a at 3000-4000 m and 0.205, 0.218 and 0.303 °C/10a at 0-1000 m, respectively. However, in spring, the rate of increase in low-elevation regions were higher than in high-elevation regions, e.g., 0.369 °C/10a at 0-1000 m and 0.317 °C/10a at 3000-4000 m.

HTD: Heterogeneous Task Decoupling for Two-Stage Object Detection.

Decoupling the sibling head has recently shown great potential in relieving the inherent task-misalignment problem in two-stage object detectors. However, existing works design similar structures for the classification and regression, ignoring task-specific characteristics and feature demands. Besides, the shared knowledge that may benefit the two branches is neglected, leading to potential excessive decoupling and semantic inconsistency. To address these two issues, we propose Heterogeneous task decoupling (HTD) framework for object detection, which utilizes a Progressive Graph (PGraph) module and a Border-aware Adaptation (BA) module for task-decoupling. Specifically, we first devise a Semantic Feature Aggregation (SFA) module to aggregate global semantics with image-level supervision, serving as the shared knowledge for the task-decoupled framework. Then, the PGraph module performs progressive graph reasoning, including local spatial aggregation and global semantic interaction, to enhance semantic representations of region proposals for classification. The proposed BA module integrates multi-level features adaptively, focusing on the low-level border activation to obtain representations with spatial and border perception for regression. Finally, we utilize the aggregated knowledge from SFA to keep the instance-level semantic consistency (ISC) of decoupled frameworks. Extensive experiments demonstrate that HTD outperforms existing detection works by a large margin, and achieves single-model 50.4%AP and 33.2% APs on COCO test-dev set using ResNet-101-DCN backbone, which is the best entry among state-of-the-arts under the same configuration. Our code is available at https://github.com/CityU-AIM-Group/HTD.

Comparison of the Trueness of Fits of the Biphasic Transverse Isotropic and Kelvin Models to the Tensile Behavior of Temporomandibular Joint Disc.

This technical brief explores the validity and trueness of fit for using the transverse isotropic biphasic and Kelvin models (first and second order generalized) for characterization of the viscoelastic tensile properties of the temporomandibular joint (TMJ) discs from pigs and goats at a strain rate of 10 mm/min. We performed incremental stress-relaxation tests from 0 to 12% strain, in 4% strain steps on pig TMJ disc samples. In addition, to compare the outcomes of these models between species, we also performed a single-step stress-relaxation test of 10% strain. The transverse isotropic biphasic model yielded reliable fits in reference to the least root mean squared error method only at low strain, while the Kelvin models yielded good fits at both low and high strain, with the second order generalized Kelvin model yielding the best fit. When comparing pig to goat TMJ disc in 10% strain stress-relaxation test, unlike the other two Kelvin models, the transverse isotropic model did not fit well for this larger step. In conclusion, the second order Kelvin model showed the best fits to the experimental data of both species. The transverse isotropic biphasic model did not fit well with the experimental data, although better at low strain, suggesting that the assumption of water flow only applies while uncrimping the collagen fibers. Thus, it is likely that the permeability from the biphasic model is not truly representative, and other biphasic models, such as the poroviscoelastic model, would likely yield more meaningful outputs and should be explored in future works.

Micro-Nanopore Structure and Fractal Characteristics of Tight Sandstone Gas Reservoirs in the Eastern Ordos Basin, China.

The complex pore system in tight sandstone reservoirs controls the storage and transport of natural gas. Thus, quantitatively characterizing the micro-nanopore structure of tight sandstone reservoirs is of great significance to determining the accumulation and distribution of tight gas. The pore structure of reservoirs was determined through polarizing microscopy, scanning electron microscopy (SEM), and the combination of mercury injection capillary pressure (MICP) and nuclear magnetic resonance (NMR) experiments on Late Paleozoic conventional and tight sandstone samples from the Linxing Block, Ordos Basin. The results show that in contrast to conventional sandstone, dissolution pores, with diameters less than 8 µm, are the main contributors to the gas storage space of tight sandstone reservoirs. The pore size distribution derived from the MICP experiment demonstrates that the main peak of tight sandstones corresponds to a pore radius in the range of 247 nm to 371 nm, while the secondary peak usually corresponds to 18 nm. The results of the NMR test illustrate that the T2 spectra of tight sandstones are unimodal, bimodal and multimodal, and the main NMR peak is highly related to the MICP peak. Fractal theory was proposed to quantitatively characterize the complex pore structure and rough porous surface. The sandstones show fractal characteristics including nanopore fractal dimension DN obtained from the MICP and large pore fractal dimension DL obtained from the NMR experiment. Both DN and DL are positively correlated with porosity and negatively correlated with permeability, demonstrating that complex and heterogeneous pore structure could increase the gas storage space and reduce the connectivity.

Micro-Nanoscale Characteristics of Pyrite and Its Implications for Gold Mineralization: Two Cases of Gold Deposits in the Youjiang Basin and Southwestern Tianshan Mountains.

The mineralogical and compositional characteristics of gold-bearing minerals and the occurrence of gold are not only of great significance to exploring the sources of ore-forming materials and their formation mechanisms but also helpful for designing reasonable beneficiations and smelting schemes and achieving remarkable economic benefits. This paper presents an integrated study on the crystal characteristics, elemental composition and distribution of pyrite (the main gold-bearing minerals), on the basis of electron probe microanalysis (EPMA), scanning electron microscopy (SEM), laser ablation-inductively coupled plasma-mass spectrometry (LA-ICP-MS) and nano-secondary ion mass spectrometry (NanoSIMS). The occurrence of gold in the Shuiyindong gold deposit and Ashawayi gold deposit has been studied by means of microscopy, SEM, and EPMA images, elemental correlations, S-Fe-As ternary diagrams, logAs-logAu diagrams and Au/As ratios. The gold in pyrite of the Shuiyindong deposit is in the form of nano gold inclusions and lattice gold. The gold in pyrite of the Ashawayi deposit dominantly exists in the form of nano gold inclusions or is present as micro-nano gold particles in the cracks or edges of pyrite, some of which can exist as lattice gold. The ore-forming hydrothermal solution of the Shuiyindong gold deposit is mainly underground hot brine, but it may be reformed by a deep magmatic hydrothermal solution or volcanic-subvolcanic hydrothermal solution. The ore-forming hydrothermal solution of the Ashawayi gold deposit is mainly derived from the metamorphic hydrothermal solution formed during the orogenic process, and the ore-forming process or post-mineralization process may be reformed by the leaching of underground hot brine. Finally, the characteristics of ore-forming fluids and evolution of the two types of deposits are determined via pyrite element surface scanning. This paper shows that micro-nanoscale study of gold-bearing pyrite is of great significance to understanding the gold mineralization process and is worth further study.

Investigation on the Structure and Fractal Characteristics of Nanopores in High-Rank Coal: Implications for the Methane Adsorption Capacity.

The structure and fractal characteristics of nanopores of high-rank coal were investigated using an approach that integrates N2 adsorption and field emission scanning electron microscopy (FE-SEM). The results indicated that the high-rank coal of the Shanxi Formation has a complex pore-fracture network composed of organic matter pores, mineral-related pores, and microfractures. The pore type of high-rank coal tends to be complicated, and the main pore types are inkbottle pores and open pores, which are more conducive to methane enrichment. The Ro,max has a negative relationship with the total pore volume. In addition, the ash and inertinite contents show a positive correlation with the average pore size (APS), while the fixed carbon content exhibits a negative relationship with the APS. The pore structure of high-rank coal is controlled not only by the degree of metamorphism but also by coal composition, which leads to the variation in pore structure becoming more complicated. With the increase in coal metamorphism, high-rank coal with high amounts of fixed carbon content generally possesses a higher irregularity in pore structure. No obvious relationship was observed between D2 and the coal components, which indicates that the pore structure, ash content, moisture content and other factors controlled by the metamorphism of coal have different effects on D2 that lead to this inapparent relationship. A negative relationship exists between adsorption volume and D1, which indicates that the high irregularity of the pore structure is not conducive to methane absorption and that no obvious correlation exists between the adsorption volume and D2. In the high-rank coal, the high D1 value represents the complexity and heterogeneity of the pore structure and represents a low adsorption affinity for methane molecules; in addition, D2 has no effect on the methane adsorption capacity.

Effect of Composition on the Micropore Structure of Non-Marine Coal-Bearing Shale: A Case Study of Permian Strata in the Qinshui Basin, China.

The nanopore network in organic-rich shale plays a key role in shale gas storage and migration, and micropores are an important structural unit in connecting the migration channel. In this study, we selected six non-marine coal-bearing shales from the Qinshui Basin to investigate the effect of composition on micropore structure using X-ray diffraction, total organic carbon (TOC), vitrinite reflectance, and CO2 adsorption methods. The results indicate that non-marine shale with higher TOC content possesses more micropores, leading to a more complex pore structure and improving the heterogeneity of shale reservoirs. With the increase in TOC content, the micropore surface area and micropore volume clearly increases, which greatly improves the gas storage space in shale reservoirs. The thermal evolution of organic matter promotes the development of micropores to a certain extent in non-marine shale. Clay minerals possess a rough surface and develop more micropores, and their influence on the micropore structure of non-marine shale is relatively strong, while terrestrial quartz exhibits significant micropore development. The obviously positive correlations between micropore volume and kaolinite, chlorite contents in the non-marine shale suggest that kaolinite and chlorite make a certain contribution to micropore volume. The characteristics of micropore structures in coal mainly depend on lithotypes, TOC content, and ash content, while clay content, quartz content, and TOC content are the key factors controlling the formation of micropores in non-marine shale.

Reaction mechanism of chloramphenicol with hydroxyl radicals for advanced oxidation processes using DFT calculations.

The structure properties of chloramphenicol (CAP), including bond information and the Fukui function for the atoms in the main chain, were investigated computationally by density functional theory (DFT). The result shows that the chiral carbons in CAP offer the most active positions for chemical reactions, which is in good agreement with the experiment. The detailed degradation mechanism for CAP with hydroxyl radicals in advanced oxidation processes is further studied at the SMD/M06-2X/6-311 + G(d,p) level of theory. The main reaction methods, including the addition-elimination reaction, hydrogen abstract reaction, hydroxyl radical addition, and bond-breaking processes, are calculated. The results show that the nitro-elimination reaction is the most likely reaction in the first step of the degradation of CAP, and the latter two processes are more likely to be hydrogen abstract reactions. The details for the transition states, intermediate radicals, and free energy surfaces for all proposed reactions are given, which makes up for a lack of experimental knowledge.

Enhanced heterogeneous Fenton-like degradation of nuclear-grade cationic exchange resin by nanoscale zero-valent iron: experiments and DFT calculations.

Nanoscale zero-valent iron (nZVI) was prepared and used as a heterogeneous Fenton-like catalyst for the degradation of nuclear-grade cationic exchange resin. The properties of nZVI before and after reaction were characterized by scanning electron microscopy (SEM), X-ray diffraction (XRD), and Brunauer-Emmett-Teller (BET) surface area analysis. The results showed that nZVI-H2O2 system exhibited the enhanced degradation of cationic resins, compared with Fe2+-H2O2, Cu0-H2O2, and Fe0/Cu0-H2O2 systems. The effects of initial temperature, nZVI dose, and H2O2 concentration were studied, and the higher temperature and nZVI dose with relatively low H2O2 concentration brought faster degradation rate. The degradation of cationic resins followed the pseudo-first-order kinetics with the apparent activation energy of 53.29 kJ/mol. According to the experimental and calculated infrared and UV-visible spectra, the carbon skeleton of cationic resins was broken with the detachment of benzene ring and the desulfonation of resin polymer by hydroxyl radicals (•OH), generating long-chain alkenes. These intermediates were further oxidized through the hydroxyl substitution, hydrogen abstraction, ring cleavage, or carbonylation reactions, finally forming carboxylic acids remained in solution.

A Statistical Model and DFT Study of the Fragmentation Mechanisms of Metronidazole by Advanced Oxidation Processes.

The degradation pathway of the antibiotic metronidazole (MNZ) in wastewater was investigated computationally with a physical statistical method and a quantum chemical approach. In both cases, density functional theory (DFT) at the M06-2X/6-311+G(d,p) level was used to calculate the structures and property parameters of all molecules. On one hand, decay of the isolated MNZ molecule excited at a given excitation energy was studied using the statistical molecular fragmentation (SMF) model. On the other hand, the reaction mechanisms of MNZ oxidized by hydroxyl radicals (•OH) in advanced oxidation processes (AOPs) were analyzed. Both studies show that the main reaction sites in MNZ are, by decreasing importance, -NO2, -CH2OH, and -CH2CH2OH. The main degradation reactions are (i) alcohol group oxidation including the abstraction of hydrogen on C in the -CH2OH group and oxidation of the hydroxyl group to the aldehyde and further to the carboxylic acid and (ii) addition-elimination reactions happening on the imidazole ring which finally replace the nitro by hydroxyl radicals. The results gained are in a good agreement with the available experimental data on MNZ degradation by AOPs. The structures of intermediates, transition states, and free energy surfaces are helpful in elucidating the details of the elimination mechanism, supplementing current experimental knowledge.

Three-dimensional macroporous graphene-wrapped zero-valent copper nanoparticles as efficient micro-electrolysis-promoted Fenton-like catalysts for metronidazole removal.

Three-dimensional macroporous graphene-wrapped zero-valent copper nanoparticles (3D-GN@Cu0) were synthesized using a self-assembly process of liquid-phase reduction and characterized by field emission scanning electron microscopy, nitrogen adsorption/desorption isotherms, X-ray diffraction, Raman spectrum analysis, and X-ray photoelectron spectroscopy. The catalytic activity of 3D-GN@Cu0 was evaluated in view of the effects of various systems, the pH value, catalyst dosage, initial metronidazole concentration and temperature, and it showed a high efficiency for removing metronidazole with saturated dissolved oxygen (without adding extra H2O2) in a wide range of pH value from 3.2 to 9.8. Combined with the results of dissolved oxygen activation, determination of reactive oxidizing species, and X-ray photoelectron spectroscopy (XPS) analysis, the surface-bounded ·OHads formed by the reaction of the in situ generation H2O2 with 3D-GN@Cu0 was mainly responsible for the removal of metronidazole. The charge distribution and electrostatic potential (ESP) of 3D-GN@Cu0 further illustrated the distribution and transfer of electrons on the catalyst surface, which predicted a micro-electrolysis-promoted Fenton-like reaction mechanism.