Performance Of Transfer Learning Research Articles

Transfer learning enables identification of multiple types of RNA modifications using nanopore direct RNA sequencing

Nanopore direct RNA sequencing (DRS) has emerged as a powerful tool for RNA modification identification. However, concurrently detecting multiple types of modifications in a single DRS sample remains a challenge. Here, we develop TandemMod, a transferable deep learning framework capable of detecting multiple types of RNA modifications in single DRS data. To train high-performance TandemMod models, we generate in vitro epitranscriptome datasets from cDNA libraries, containing thousands of transcripts labeled with various types of RNA modifications. We validate the performance of TandemMod on both in vitro transcripts and in vivo human cell lines, confirming its high accuracy for profiling m6A and m5C modification sites. Furthermore, we perform transfer learning for identifying other modifications such as m7G, Ψ, and inosine, significantly reducing training data size and running time without compromising performance. Finally, we apply TandemMod to identify 3 types of RNA modifications in rice grown in different environments, demonstrating its applicability across species and conditions. In summary, we provide a resource with ground-truth labels that can serve as benchmark datasets for nanopore-based modification identification methods, and TandemMod for identifying diverse RNA modifications using a single DRS sample.

Physics-Informed Transfer Learning to Enhance Sleep Staging.

At-home sleep staging using wearable medical sensors poses a viable alternative to in-hospital polysomnography due to its lower cost and lower disruption to the daily lives of patients, especially in the case of long-term monitoring. Machine learning with wearables however is difficult due to the paucity of data from wearable sensors, making automation a challenge. Transfer learning from hospital polysomnograms can boost performance, but is still hindered by differences between wearable and in-hospital EEG resulting in part from differing electrode placement. We improve transfer learning performance by using electrophysiological models of a human head to generate synthetic EEG resembling EEG from a wearable sensor. The data generation method utilizes Low-Resolution Electromagnetic Tomography Analysis (LORETA). Real EEG from standard in- hospital recordings is first mapped to point currents within the brain using LORETA, after which the point currents are used to estimate EEG that would have been recorded using a wearable sensor at any given point on the head. Augmenting source datasets with synthetic data statistically significantly boosted accuracy on a wearable sleep staging task from 80.8% to 81.3% on average, depending on the transfer learning parameters and data sources. Machine learning performance can be improved using data synthesized using physical models. Our approach represents a new form of transfer learning and demonstrates that incorporating domain knowledge of electrophysiological modeling can improve machine learning results for sleep staging tasks. We expect this approach to be particularly useful for EEG data which is hard to collect, or which is obtained using unusual electrode configurations.

Learning-Based Approach for Automated Surface Inspection with Industrial Tomography Imaging

Abstract In recent years, advanced deep learning techniques have emerged as pivotal tools in enabling the development of robust vision-based solutions for steel surface inspection. This resulted in enhanced inspection accuracy, all while significantly reducing costs in the manufacturing industry. However, the lack of actual steel surface defects datasets currently places a certain constraint on further research into classifying those anomalies. As a consequence, the Convolutional Neural Network (CNN) technique, known for its prowess in image-related tasks, faces certain challenges especially in classifying less common defects. This work proposes a novel hybrid CNN model with an SVM classifier at the output layer for surface defects classification. The features extracted from the pre-trained ResNet152 and EfficientB0 CNN algorithms are concatenated and fed to the SVM layer for classification. Extensive experiments on a merged dataset consisting of the publicly available Northeastern University (NEU) dataset and Xsteel surface defect dataset (X-SDD) are carried out and the accuracy and F1 scores are calculated for performance evaluation. The merged dataset contains eleven typical defect types with a total of 2660 defect images. Then, the adopted algorithm is compared with ten fine-tuned deep learning models to evaluate the performance of transfer learning for steel defect detection and identification. The evaluation results show that the deep feature extraction and SVM classification produced better results than the transfer learning. Finally, the proposed classifier model is validated on a new collected dataset from a Computed Tomography scanner with an accuracy reaching over 96%.

An LSTM-based adversarial variational autoencoder framework for self-supervised neural decoding of behavioral choices.

Objective. This paper presents data-driven solutions to address two challenges in the problem of linking neural data and behavior: 1) unsupervised analysis of behavioral data and automatic label generation from behavioral observations, and 2) extraction of subject-invariant features for the development of generalized neural decoding models.

Approach. For behavioral analysis and label generation, an unsupervised method, which employs an autoencoder to transform behavior data into a cluster-friendly feature space is presented. The model iteratively refines the assigned clusters with soft clustering assignment loss, and gradually improves the learned feature representations. To address subject variability in decoding neural activity, adversarial learning in combination with a long short-term memory-based adversarial variational autoencoder (LSTM-AVAE) model is employed. By using an adversary network to constrain the latent representations, the model captures shared information among subjects' neural activity, making it proper for cross-subject transfer learning.

Main results. The proposed approach is evaluated using cortical recordings of Thy1-GCaMP6s transgenic mice obtained via widefield calcium imaging during a motivational licking behavioral experiment. The results show that the proposed model achieves an accuracy of 89.7% in cross-subject neural decoding, outperforming other well-known autoencoder-based feature learning models. These findings suggest that incorporating an adversary network eliminates subject dependency in representations, leading to improved cross-subject transfer learning performance, while also demonstrating the effectiveness of LSTM-based models in capturing the temporal dependencies within neural data.

Significance. Results demonstrate the feasibility of the proposed framework in unsupervised clustering and label generation of behavioral data, as well as achieving high accuracy in cross-subject neural decoding, indicating its potentials for relating neural activity to behavior.

Self-supervised contrastive learning for EEG-based cross-subject motor imagery recognition.

Objective. The extensive application of electroencephalography (EEG) in brain-computer interfaces (BCIs) can be attributed to its non-invasive nature and capability to offer high-resolution data. The acquisition of EEG signals is a straightforward process, but the datasets associated with these signals frequently exhibit data scarcity and require substantial resources for proper labeling. Furthermore, there is a significant limitation in the generalization performance of EEG models due to the substantial inter-individual variability observed in EEG signals.Approach. To address these issues, we propose a novel self-supervised contrastive learning framework for decoding motor imagery (MI) signals in cross-subject scenarios. Specifically, we design an encoder combining convolutional neural network and attention mechanism. In the contrastive learning training stage, the network undergoes training with the pretext task of data augmentation to minimize the distance between pairs of homologous transformations while simultaneously maximizing the distance between pairs of heterologous transformations. It enhances the amount of data utilized for training and improves the network's ability to extract deep features from original signals without relying on the true labels of the data.Main results. To evaluate our framework's efficacy, we conduct extensive experiments on three public MI datasets: BCI IV IIa, BCI IV IIb, and HGD datasets. The proposed method achieves cross-subject classification accuracies of 67.32%, 82.34%, and 81.13%on the three datasets, demonstrating superior performance compared to existing methods.Significance. Therefore, this method has great promise for improving the performance of cross-subject transfer learning in MI-based BCI systems.

CT-based intratumoral and peritumoral deep transfer learning features prediction of lymph node metastasis in non-small cell lung cancer.

The main metastatic route for lung cancer is lymph node metastasis, and studies have shown that non-small cell lung cancer (NSCLC) has a high risk of lymph node infiltration. This study aimed to compare the performance of handcrafted radiomics (HR) features and deep transfer learning (DTL) features in Computed Tomography (CT) of intratumoral and peritumoral regions in predicting the metastatic status of NSCLC lymph nodes in different machine learning classifier models. We retrospectively collected data of 199 patients with pathologically confirmed NSCLC. All patients were divided into training (n = 159) and validation (n = 40) cohorts, respectively. The best HR and DTL features in the intratumoral and peritumoral regions were extracted and selected, respectively. Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Light Gradient Boosting Machine (Light GBM), Multilayer Perceptron (MLP), and Logistic Regression (LR) models were constructed, and the performance of the models was evaluated. Among the five models in the training and validation cohorts, the LR classifier model performed best in terms of HR and DTL features. The AUCs of the training cohort were 0.841 (95% CI: 0.776-0.907) and 0.955 (95% CI: 0.926-0.983), and the AUCs of the validation cohort were 0.812 (95% CI: 0.677-0.948) and 0.893 (95% CI: 0.795-0.991), respectively. The DTL signature was superior to the handcrafted radiomics signature. Compared with the radiomics signature, the DTL signature constructed based on intratumoral and peritumoral areas in CT can better predict NSCLC lymph node metastasis.

SegIns: A simple extension to instance discrimination task for better localization learning

Recent self-supervised learning methods, where instance discrimination task is a fundamental way of pretraining convolutional neural networks (CNN), excel in transfer learning performance. Even though instance discrimination task is a well suited pretraining method for classification with its image-level learning, lack of dense representation learning makes it sub-optimal for localization tasks such as object detection. In this paper, we aim to mitigate this shortcoming of instance discrimination task by extending it to jointly learn dense representations alongside image-level representations. We add a segmentation branch parallel to the image-level learning to predict class-agnostic masks, enhancing location-awareness of the representations. We show the effectiveness of our pretraining approach on localization tasks by transferring the learned representations to object detection and segmentation tasks, providing relative improvements by up to 1.7% AP on PASCAL VOC and 0.8% AP on COCO object detection, 0.8% AP on COCO instance segmentation and 3.6% mIoU on PASCAL VOC semantic segmentation respectively.

Improved Adversarial Transfer Network for Bearing Fault Diagnosis under Variable Working Conditions

Bearings are one of the critical components of rotating machinery, and their failure can cause catastrophic consequences. In this regard, previous studies have proposed a variety of intelligent diagnosis methods. Most existing bearing fault diagnosis methods implicitly assume that the training and test sets are from the same distribution. However, in real scenarios, bearings have been working in complex and changeable working environments for a long time. The data during their working processes and the data used for model training cannot meet this condition. This paper proposes an improved adversarial transfer network for fault diagnosis under variable working conditions. Specifically, this paper combines an adversarial transfer network with a short-time Fourier transform to obtain satisfactory results with the lighter network. Then, this paper employs a channel attention module to enhance feature fusion. Moreover, this paper designs a novel domain discrepancy hybrid metric loss to improve model transfer learning performance. Finally, this paper verifies the method’s effectiveness on three datasets, including dual-rotor, a Case Western Reserve University dataset and the Ottawa dataset. The proposed method achieves average accuracy, surpassing other methods, and shows better domain alignment capabilities.

Procedural Content Generation via Knowledge Transformation (PCG-KT)

We introduce the concept of Procedural Content Generation via Knowledge Transformation (PCG-KT), a new lens and framework for characterizing PCG methods and approaches in which content generation is enabled by the process of knowledge transformation—transforming knowledge derived from one domain in order to apply it in another. Our work is motivated by a substantial number of recent PCG works that focus on generating novel content via repurposing derived knowledge. Such works have involved, for example, performing transfer learning on models trained on one game's content to adapt to another game's content, as well as recombining different generative distributions to blend the content of two or more games. Such approaches arose in part due to limitations in PCG via Machine Learning (PCGML) such as producing generative models for games lacking training data and generating content for entirely new games. In this paper, we categorize such approaches under this new lens of PCG-KT by offering a definition and framework for describing such methods and surveying existing works using this framework. Finally, we conclude by highlighting open problems and directions for future research in this area.

Variational inference-based transfer learning for profile monitoring with incomplete data

Profile monitoring is a widely used tool in quality control. The rapid development of sensor technology has created unprecedented opportunities for multi-channel profile data collection, which motivates the modeling and transfer learning for multi-profile data. The Multi-output Gaussian Process (MGP) is often used for multi-profile data, due to its flexible modeling capability and elegant mathematical properties. However, two practical concerns limit the broader application of MGP for transfer learning and monitoring of multi-profile data: high computational complexity and data incompleteness. In this article, we propose a Variational Inference (VI)-based MGP framework to facilitate transfer learning and profile monitoring using incomplete profile data. The proposed framework features a specially designed convolutional structure for constructing an explicit covariance relationship between the inducing variables in VI and the MGP in multi-profile data. This structure inspires a comprehensive solution to both computational complexity and data incompleteness in modeling multi-profile data, which facilities the transfer learning for profile monitoring. Various numerical studies and one real case study are conducted to demonstrate and compare the transfer learning and monitoring performance of the proposed method. The results show the method can achieve superior monitoring performance while maintain a very low level of computational complexity.

Fine classification of crops based on an inductive transfer learning method with compact polarimetric SAR images

ABSTRACT Compact polarimetric synthetic aperture radar (CP SAR) reduces fully polarimetric SAR system complexity and expands the imaging swath. Generally, fine classification of crop types relies on many labeled training samples. However, due to the temporal interval of crop phenology and ground environment variations over time, training samples from one dataset usually perform poorly for another. Therefore, in this study, transfer learning is introduced to crop classification to ensure classification accuracy by improving reusability of training samples. A stable and robust inductive transfer learning method, i.e. the Transfer Bagging-based Ensemble Learning (TBEL) algorithm, is proposed. The main idea is to select an adequate number of representative samples from unlabeled datasets to characterize each class in the target domain based on limited labeled samples and construct a classifier set to classify the target domain. This study investigates CP SAR data performance in transfer learning for crop classification. The proposed algorithm in the experimental study is compared with six typical methods (Subspace Alignment (SA), CORrelation ALignment (CORAL), Joint Distribution Adaptation (JDA), Balanced Distribution Adaptation (BDA), Transfer Bagging (TrBagg), and Bagging-based Ensemble Transfer Learning (BETL)). The experimental results show that the crop classification accuracy based on the TBEL algorithm is more stable, with an improved overall classification accuracy of 2–6%. Classifying the same rice harvest stage in the cross-year domain has the highest overall accuracy of 92.2%. Wheat fields in different scenes are also classified. Based on the TBEL algorithm, the overall classification accuracy improves by 1–10% compared with typical methods, with an accuracy of at least 87.6%. Furthermore, by testing the CP mode classification performance over various crops in transfer learning, we find that the circular CP mode performs better than the linear mode in most cases. This conclusion agrees with single-scene applications and was first verified in transfer learning.

Comprehensive Sensitivity Analysis Framework for Transfer Learning Performance Assessment for Time Series Forecasting: Basic Concepts and Selected Case Studies

Recently, transfer learning has gained popularity in the machine learning community. Transfer Learning (TL) has emerged as a promising paradigm that leverages knowledge learned from one or more related domains to improve prediction accuracy in a target domain with limited data. However, for time series forecasting (TSF) applications, transfer learning is relatively new. This paper addresses the need for empirical studies as identified in recent reviews advocating the need for practical guidelines for Transfer Learning approaches and method designs for time series forecasting. The main contribution of this paper is the suggestion of a comprehensive framework for Transfer Learning Sensitivity Analysis (SA) for time series forecasting. We achieve this by identifying various parameters seen from various angles of transfer learning applied to time series, aiming to uncover factors and insights that influence the performance of transfer learning in time series forecasting. Undoubtedly, symmetry appears to be a core aspect in the consideration of these factors and insights. A further contribution is the introduction of four TL performance metrics encompassed in our framework. These TL performance metrics provide insight into the extent of the transferability between the source and the target domains. Analyzing whether the benefits of transferred knowledge are equally or unequally accessible and applicable across different domains or tasks speaks to the requirement of symmetry or asymmetry in transfer learning. Moreover, these TL performance metrics inform on the possibility of the occurrence of negative transfers and also provide insight into the possible vulnerability of the network to catastrophic forgetting. Finally, we discuss a sensitivity analysis of an Ensemble TL technique use case (with Multilayer Perceptron models) as a proof of concept to validate the suggested framework. While the results from the experiments offer empirical insights into various parameters that impact the transfer learning gain, they also raise the question of network dimensioning requirements when designing, specifically, a neural network for transfer learning.

NanoCon: contrastive learning-based deep hybrid network for nanopore methylation detection.

5-Methylcytosine (5mC), a fundamental element of DNA methylation in eukaryotes, plays a vital role in gene expression regulation, embryonic development, and other biological processes. Although several computational methods have been proposed for detecting the base modifications in DNA like 5mC sites from Nanopore sequencing data, they face challenges including sensitivity to noise, and ignoring the imbalanced distribution of methylation sites in real-world scenarios. Here, we develop NanoCon, a deep hybrid network coupled with contrastive learning strategy to detect 5mC methylation sites from Nanopore reads. In particular, we adopted a contrastive learning module to alleviate the issues caused by imbalanced data distribution in nanopore sequencing, offering a more accurate and robust detection of 5mC sites. Evaluation results demonstrate that NanoCon outperforms existing methods, highlighting its potential as a valuable tool in genomic sequencing and methylation prediction. In addition, we also verified the effectiveness of our representation learning ability on two datasets by visualizing the dimension reduction of the features of methylation and nonmethylation sites from our NanoCon. Furthermore, cross-species and cross-5mC methylation motifs experiments indicated the robustness and the ability to perform transfer learning of our model. We hope this work can contribute to the community by providing a powerful and reliable solution for 5mC site detection in genomic studies. The project code is available at https://github.com/Challis-yin/NanoCon.

On the relation between K–L divergence and transfer learning performance on causality extraction tasks

The problem of extracting causal relations from text remains a challenging task, even in the age of Large Language Models (LLMs). A key factor that impedes the progress of this research is the availability of the annotated data and the lack of common labeling methods. We investigate the applicability of transfer learning (domain adaptation) to address these impediments in experiments with three publicly available datasets: FinCausal, SCITE, and Organizational. We perform pairwise transfer experiments between the datasets using DistilBERT, BERT, and SpanBERT (variants of BERT) and measure the performance of the resulting models. To understand the relationship between datasets and performance, we measure the differences between vocabulary distributions in the datasets using four methods: Kullback–Leibler (K–L) divergence, Wasserstein metric, Maximum Mean Discrepancy, and Kolmogorov–Smirnov test. We also estimate the predictive capability of each method using linear regression. We record the predictive values of each measure. Our results show that K–L divergence between the distribution of the vocabularies in the data predicts the performance of the transfer learning with R2 = 0.0746. Surprisingly, the Wasserstein distance predictive value is low (R2=0.52912), and the same for the Kolmogorov–Smirnov test (R2 =0.40025979). This is confirmed in a series of experiments. For example, with variants of BERT, we observe an almost a 29% to 32% increase in the macro-average F1-score, when the gap between the training and test distributions is small, according to the K–L divergence — the best-performing predictor on this task. We also discuss these results in the context of the sub-par performance of some large language models on causality extraction tasks. Finally, we report the results of transfer learning informed by K–L divergence; namely, we show that there is a 12 to 63% increase in the performance when a small portion of the test data is added to the training data. This shows that corpus expansion and n-shot learning benefit, when the process of choosing examples maximizes their information content, according to the K–L divergence.

Demonstration of transfer learning using 14 nm technology analog ReRAM array

Analog memory presents a promising solution in the face of the growing demand for energy-efficient artificial intelligence (AI) at the edge. In this study, we demonstrate efficient deep neural network transfer learning utilizing hardware and algorithm co-optimization in an analog resistive random-access memory (ReRAM) array. For the first time, we illustrate that in open-loop deep neural network (DNN) transfer learning for image classification tasks, convergence rates can be accelerated by approximately 3.5 times through the utilization of co-optimized analog ReRAM hardware and the hardware-aware Tiki-Taka v2 (TTv2) algorithm. A simulation based on statistical 14 nm CMOS ReRAM array data provides insights into the performance of transfer learning on larger network workloads, exhibiting notable improvement over conventional training with random initialization. This study shows that analog DNN transfer learning using an optimized ReRAM array can achieve faster convergence with a smaller dataset compared to training from scratch, thus augmenting AI capability at the edge.

Performance assessment of cross office building energy prediction in the same region using the domain adversarial transfer learning strategy

Precise and reliable building energy consumption prediction is of great significance for building energy management and future energy planning. To mitigate the adverse impact of data limitation on the performance of building energy consumption prediction model, a transfer learning strategy was introduced. However, the operation modes of different building energy systems are different, resulting in a large difference in the distribution of building energy consumption related data. In this paper, the relevant data of 25 buildings from the same region of the United States were selected. On this basis, two similarity calculation strategies are compared, and three similarity measurement methods were used to analyze the impact of different source buildings on the performance improvement of transfer learning models. The results show that choosing the appropriate source building can improve the transfer learning performance by up to 82.56% when compared to the no-transfer model (the baseline model). When higher similarity source buildings are selected, the corresponding average PIR is about 24.7% higher than the total average, which can make the transfer model performance improvement more obvious. This paper provides insights into the selection of source domains for transfer learning strategies under the same region.

WECNN-PDP: Weighted Ensemble Convolutional Neural Networks Models to Improve the Plant Disease Prediction

As an agricultural country, Indonesia’s agricultural production is essential. However, crop failure will occur if diseases and other factors, such as natural disasters, attack many plant fields. These problems can be minimized by early detection of plant diseases. However, detection will be challenging if done conventionally. Prior research has shown that deep learning algorithms can perform detection with promising results. In this study, we propose a new weighted deep learning ensemble method as a solution for better performance in plant disease detection. We ensemble the model by considering the combination of two and three pre-trained convolutional neural networks (CNNs). Initially, we perform transfer learning on individual CNN models by prioritizing high-dimensional features through weight updates on the last few layers. Finally, we ensemble the models by finding the best weights for each model using grid search. Experimental results on the Plant Village dataset indicate that our model has improved the classification of 38 plant diseases. Based on metrics, the three-model ensemble performed better than the two-model ensemble. The best accuracy results of the ensemble MobileNetV2-DenseNet121 and MobileNetV2-Xception-DenseNet121 models are 99.49% and 99.56%, respectively. In addition, these models are also better than the state-of-the-art models and previous feature fusion techniques we proposed in LEMOXINET. Based on these results, the ensemble technique improved the detection performance, and it is expected to be applied to real-world conditions and can be a reference to be developed further in future research.

Transferring U-Net between low-dose CT denoising tasks: a validation study with varied spatial resolutions.

Recently, deep learning techniques have been widely used in low-dose computed tomography (LDCT) imaging applications for quickly generating high quality computed tomography (CT) images at lower radiation dose levels. The purpose of this study is to validate the reproducibility of the denoising performance of a given network that has been trained in advance across varied LDCT image datasets that are acquired from different imaging systems with different spatial resolutions. Specifically, LDCT images with comparable noise levels but having different spatial resolutions were prepared to train the U-Net. The number of CT images used for the network training, validation and test was 2,400, 300 and 300, respectively. Afterwards, self- and cross-validations among six selected spatial resolutions (62.5, 125, 250, 375, 500, 625 µm) were studied and compared side by side. The residual variance, peak signal to noise ratio (PSNR), normalized root mean square error (NRMSE) and structural similarity (SSIM) were measured and compared. In addition, network retraining on a small number of image set was performed to fine tune the performance of transfer learning among LDCT tasks with varied spatial resolutions. Results demonstrated that the U-Net trained upon LDCT images having a certain spatial resolution can effectively reduce the noise of the other LDCT images having different spatial resolutions. Regardless, results showed that image artifacts would be generated during the above cross validations. For instance, noticeable residual artifacts were presented at the margin and central areas of the object as the resolution inconsistency increased. The retraining results showed that the artifacts caused by the resolution mismatch can be greatly reduced by utilizing about only 20% of the original training data size. This quantitative improvement led to a reduction in the NRMSE from 0.1898 to 0.1263 and an increase in the SSIM from 0.7558 to 0.8036. In conclusion, artifacts would be generated when transferring the U-Net to a LDCT denoising task with different spatial resolution. To maintain the denoising performance, it is recommended to retrain the U-Net with a small amount of datasets having the same target spatial resolution.

Detection of Papilledema Severity from Color Fundus Images using Transfer Learning Approaches

Papilledema is edema in the area where the optic nerve meets the eye as a result of increased pressure inside the head. This disease can result in very serious problems, such as abnormal optical changes, decreased visual acuity, and even permanent blindness if left untreated. In this study, an image processing based solution was presented for the detection of papilledema severity from color fundus images using transfer learning approaches. The image dataset includes 295 papilledema images, 295 pseudopapilledema images, and 779 control images. Histogram equalization and the 3D box filter were used for image preprocessing. The images were enhanced with the histogram equalization method and denoised with the 3D box filter method. Then, the performances of EfficentNet-B0, GoogLeNet, MobileNetV2, NASNetMobile, and ResNet-101 transfer learning approaches were compared. The hold-out method was used to calculate the performance of transfer learning. In the experiments, the MobileNetV2 approach had the highest performance with 0.96 overall accuracy and 0.94 Cohen's Kappa. The results of the experiments proved that the combination of the histogram equalization, the 3D box filter, and the MobileNetV2 transfer learning approach can be used for automatic detection of papilledema severity. Compared to other similar studies that are known in the literature, the overall accuracy was higher.

Gastro-BaseNet: A Specialized Pre-Trained Model for Enhanced Gastroscopic Data Classification and Diagnosis of Gastric Cancer and Ulcer.

Most of the development of gastric disease prediction models has utilized pre-trained models from natural data, such as ImageNet, which lack knowledge of medical domains. This study proposes Gastro-BaseNet, a classification model trained using gastroscopic image data for abnormal gastric lesions. To prove performance, we compared transfer-learning based on two pre-trained models (Gastro-BaseNet and ImageNet) and two training methods (freeze and fine-tune modes). The effectiveness was verified in terms of classification at the image-level and patient-level, as well as the localization performance of lesions. The development of Gastro-BaseNet had demonstrated superior transfer learning performance compared to random weight settings in ImageNet. When developing a model for predicting the diagnosis of gastric cancer and gastric ulcers, the transfer-learned model based on Gastro-BaseNet outperformed that based on ImageNet. Furthermore, the model's performance was highest when fine-tuning the entire layer in the fine-tune mode. Additionally, the trained model was based on Gastro-BaseNet, which showed higher localization performance, which confirmed its accurate detection and classification of lesions in specific locations. This study represents a notable advancement in the development of image analysis models within the medical field, resulting in improved diagnostic predictive accuracy and aiding in making more informed clinical decisions in gastrointestinal endoscopy.