MSMoch Arief Soeleman

Moch Arief Soeleman

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Moch Arief Soeleman has a total of 96 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Mixed-source-to-institutional hospital-holdout evaluation of DICOM-aware adaptive harmonization for chest radiograph lesion detection

    by Abu Salam, Cinantya Paramita, Pulung Nurtantio Andono, Purwanto, Moch Arief Soeleman, Erwin Yudi Hidayat, Defri Kurniawan, Daniel Rizky, Budi Setiawan, Antonius Gunawan Santoso, et al.

  • Open Access

    Hospital-Held-Out Workflow-Level Evaluation of DICOM-Aware Preprocessing for Evidence-Grounded Chest Radiograph Finding Support

    by Abu Salam, Danang Wahyu Utomo, Pulung Nortantiyo Andono, Purwanto Purwanto, Moch Arief Soeleman, Cinantya Paramita, Sindhu Rakasiwi, Yani Parti Astuti, Egia Rosi Subhiyakto, Mohamad Sidiq, et al.

    Purpose: To evaluate whether DICOM-aware adaptive HyFusion improves an evidence-grounded chest radiograph (CXR) finding-support workflow compared with a baseline DICOM-derived image branch.Methods: This retrospective multi-source public–private pilot study used institutional CXR data together with NIH, MS-CXR, and VinDr-CXR data. A five-class grounding ontology was used for Atelectasis, Consolidation, Pleural Effusion, Pneumothorax, and Parenchymal Opacity/Infiltration. DICOM-derived images underwent metadata-aware preprocessing, including photometric polarity correction, before paired baseline and adaptive HyFusion branches were generated. A Faster R-CNN detector extracted class-specific bounding-box evidence. Detector predictions were converted into thresholded evidence tables and used by a controlled grounded finding-support module to generate evidence-mediated finding statements. The primary endpoint was a private hospital-held-out locked test set, and the primary finding-support analysis used a fixed evidence threshold of 0.30.Results: The final polarity-corrected experiment included 4,454 CXR images. On the private hospital-held-out test set, HyFusion increased detector mAP from 0.0098 to 0.0172 and reduced false positives per negative image from 6.68 to 5.44, although mAP50 changed from 0.0452 to 0.0416. At the finding-support level, HyFusion improved finding F1 from 0.235 to 0.333 and recall from 0.314 to 0.490. Hallucinated finding rate decreased from 0.812 to 0.747, and missing finding rate decreased from 0.686 to 0.510.Conclusion: In this pilot detector-to-report workflow, DICOM-aware adaptive HyFusion improved downstream evidence-grounded finding-support performance on the primary private hospital-held-out test set after polarity-aware preprocessing. These findings support workflow-level evaluation of medical imaging AI preprocessing methods rather than relying on detector metrics alone.

  • Open Access

    Investigating Cross-Hospital Generalization of Chest Radiograph Lesion Detection Through Public-to-Institutional Source-Holdout Evaluation

    by Abu Salam, Cinantya Paramita, Pulung Nortantiyo Andono, Purwanto Purwanto, Moch Arief Soeleman, Erwin Yudi Hidayat, Defri Kurniawan, Daniel Rizky, Budi Setiawan, Antonius Gunawan Santoso, et al.

    Robust cross-hospital generalization remains a major challenge in chest radiograph lesion detection, especially when detectors trained on mixed public and institutional sources are transferred to an unseen hospital. This study investigated whether a Digital Imaging and Communications in Medicine (DICOM)-aware adaptive harmonization branch, Adaptive HyFusion, improves public-to-institutional source-holdout transfer compared with standard preprocessing. A fixed Faster R-CNN detector with a ResNet-50-FPN backbone was evaluated under a harmonized five-class lesion ontology using VinDr-CXR, MS-CXR, and two non-public Indonesian hospital cohorts. Scenario 2A, using RSUD Dr. Adhyatma, MPH Jawa Tengah for institutional development and RSUD Sunan Kalijaga Kabupaten Demak as the locked held-out test source, was the main analysis; Scenario 2B reversed the hospital roles as a sensitivity analysis. Performance was assessed using held-out mean average precision at an intersection-over-union threshold of 0.50 (mAP50), class-wise AP50, recall, precision, false positives per image, and validation-to-held-out degradation. In Scenario 2A, Adaptive HyFusion achieved comparable held-out mAP50 to baseline preprocessing (0.1169 ± 0.0460 vs. 0.1156 ± 0.0545), with higher recall (0.1282 ± 0.0513 vs. 0.1026 ± 0.0128), higher precision (0.0924 ± 0.0053 vs. 0.0728 ± 0.0143), and lower false positives per image (1.2215 ± 0.3754 vs. 1.2673 ± 0.1866). In Scenario 2B, Adaptive HyFusion achieved higher held-out mAP50 (0.1236 ± 0.0377 vs. 0.1070 ± 0.0102), higher precision, lower false-positive burden, and smaller validation-to-held-out mAP50 degradation, although recall was slightly lower. Class-wise effects remained heterogeneous. These findings support DICOM-aware adaptive preprocessing as a meaningful component of cross-hospital generalization strategy rather than a universally superior detection solution.

  • Open Access

    A reliability-aware analytics framework for multi-vendor Tuberculosis screening from chest radiographs

    by Abu Salam, Pulung Nurtantio Andono, Purwanto, Moch Arief Soeleman, Antonius Gunawan Santoso, Ika Novita Dewi, Sasha-Gay Wright

  • Open Access

    Hybrid SAR-Optical Remote Sensing for Flood Inundation Mapping: Feature Contribution Analysis in a Wetland Environment

    by Desy Ika Puspitasari, Edi Noersasongko, Purwanto, Moch Arief Soeleman, Catur Supriyanto

    Flooding is a recurrent problem in Banjar Regency due to its low-lying wetland topography, high rainfall, andriver sedimentation. Nonetheless, precise flood mapping continues to be a challenge due to the interference of cloud coveron optical imaging and spectral ambiguity in diverse wetland ecosystems. To address these limitations, this study proposesa hybrid SAR–optical method that combines Sentinel-1 backscatter with Sentinel-2 water indices (FWEI and AWEI) usingpixel-level feature stacking and feature contribution analysis. Four machine learning classifiers—Random Forest, LogisticRegression, Support Vector Machine (SVM), and XGBoost—were evaluated using a spatially independent validation schemeto ensure robust generalization. The results indicate that AWEI surpasses individual features in water discrimination, whereasthe combination of SAR and AWEI consistently enhances classification stability and spatial coherence. Among the assessedmodels, SVM and RF exhibit similar performance, with SVM attaining the optimal classification balance (OA = 0.98, Kappa= 0.97, F1 = 0.98) for water detection, while the difference lacks statistical significance. Flood inundation maps were derivedthrough change detection between independently classified pre- and post-flood water maps. The results indicate that multisensor integration enhances the delineation of flood-affected areas, particularly in complex wetland environments. However,the study also highlights that the reliability of flood inundation mapping is inherently dependent on the accuracy of pre- andpost-flood water classification as well as the temporal consistency of the input imagery. Furthermore, feature contributionanalysis combined with spatial validation reveals that AWEI provided the strongest standalone classification performance,while SAR contributed complementary structural information that improved classification robustness. These findingsdemonstrate that spatial validation effectively mitigates autocorrelation bias, leading to more reliable and generalizablewater classification performance for flood inundation analysis.

  • Open Access

    HyFusion-v2: DICOM-aware hybrid normalization pipeline with QC-guarded adaptive fusion for cross-manufacturer chest radiographs

    by Abu Salam, Pulung Nortantiyo Andono, Purwanto Purwanto, Moch Arief Soeleman, Antonius Gunawan Santoso, Ika Novita Dewi, Farid Che Ghazali

    Chest radiograph datasets aggregated across manufacturers exhibit variability in pixel encoding, spatial resolution, and intensity representation, introducing technical domain heterogeneity that challenges reproducible and vendor-robust preprocessing. We present HyFusion-v2, a DICOM-aware hybrid normalization pipeline for cross-manufacturer harmonization with an explicit quality-control safeguard. The method standardizes DICOM interpretation and image geometry, then performs domain-adaptive hybrid normalization guided by a Domain Heterogeneity Index (DHI) that quantifies manufacturer-level acquisition variability using metadata and early-decoded intensity statistics. QC-guarded step-down adjustment and best-candidate fallback mechanism evaluates structural preservation and contrast stability to constrain over-enhancement. HyFusion-v2 was evaluated on 3,502 retrospective DICOM chest radiographs acquired from two manufacturers. The method achieved a balanced structural–radiometric trade-off (median SSIM 0.93, PSNR 30.80 dB, CNR 23.24), with adaptive adjustments/fallback triggered in 6.8–15.7% of cases to maintain predefined quality thresholds. No downstream model training or clinical performance evaluation was performed. The method:integrates DICOM-aware standardization with hybrid normalization for vendor-robust preprocessing;enables domain-level adaptation through DHI-guided fusion without requiring model retraining;incorporates objective QC with automatic rollback to ensure conservative harmonization.

  • Open Access

    The Impact of Squeeze-and-Excitation Blocks on CNN Models and Transfer Learning for Pneumonia Classification Using Chest X-ray Images

    by Muhammad Yunan, Aris Marjuni, Affandy Affandy, Mochamad Arief Soeleman, Iqbal Firdaus

    Pneumonia is one of the leading causes of death due to respiratory tract infections, especially in children and the elderly. Early detection using chest X-ray images is crucial to accelerate diagnosis and treatment, but manual interpretation is often subjective and error-prone. This study evaluates the effect of Squeeze-and-Excitation (SE) Block integration on the performance of a custom Convolutional Neural Network (CNN) model and three popular transfer learning architectures: MobileNetV2, VGG16, and InceptionV3 in X-ray image-based pneumonia classification. A dataset of 5,856 images, taken from Chest X-ray Images (Pneumonia) on Kaggle, was processed through preprocessing, undersampling, and augmentation. Each model was tested in two configurations: without and with SE Block. Evaluation was performed using accuracy, precision, recall, F1-score, and test loss metrics. The results show that SE Block integration improves the performance of most models. The accuracy of the custom CNN increased from 95.17% to 95.88%, MobileNetV2 from 97.18% to 97.59%, and VGG16 from 96.88% to 97.69%. InceptionV3 also saw an accuracy increase from 94.06% to 94.16%, although accompanied by an increase in test loss. SE Block proved effective in strengthening the model's emphasis on important features through an inter-channel recalibration mechanism, especially on efficient architectures like MobileNetV2 and complex models like VGG16. These findings support the development of a more accurate, efficient, and adaptive deep learning-based pneumonia diagnosis system, especially for implementation in healthcare facilities with limited resources.

  • Open Access

    Peningkatan Akurasi Klasifikasi Awan Cumulonimbus Dari Satelit Himawari-8 Saat Cuaca Ekstrem Dengan Menggunakan Metode Grayscale Thermal Image Dan Neural Network Backpropagation

    by Triyotomo Triyotomo, Sari Ayu Wulandari, M. Arief Soeleman, Farrikh Al Zami

    Cumulonimbus is the only type of cloud that can produce hail, lightning, and thunder. This type of cloud can cause extreme weather that causes damage to public infrastructure and can also cost lives. This research aims to improve Cumulonimbus cloud detection on the Himawari-8 satellite using a combination of the Grayscale Thermal Image method and the method of Artificial Neural Network Backpropagation. The data was taken during the transition season, which is a potential time the onset of extreme weather caused by Cumulonimbus clouds is quite large, and the consequences incurred can cause very significant losses. To detect Cumulonimbus, The Himawari-8 Satellite Image is pre-processed so that an image is obtained gray thermal, then the image is converted into digital data in the form of numbers and from the characterization of the results using histograms. The last process is classified using Artificial Neural Network Propagation. All processes in this study use Matlab to obtain the best classification accuracy. The expected result is an increase in the value of accuracy when using the method of grayscale thermal image compared without using this method. Each accuracy value training data, validating data, and testing data obtained increased from 96.6%, 84.46%, and 80.02 to 100%, 88.9%, and 91.7%.

  • Open Access

    Hyperparameter Tuning with Optuna to optimize the YOLOv11n Model for Weed Detection

    by Candhy Fadhila Arsyad, Pulung Nurtantio Andono, Moch Arief Soeleman

    Accurate weed detection is essential for maintaining the cleanliness and aesthetic appeal of residential yards. This study aimed to optimize YOLOv11n, a lightweight object detection model, to achieve high precision in weed identification under real-world conditions. The novelty of this study lies in the application of Optuna, an automatic hyperparameter optimization framework, to enhance model performance while maintaining computational efficiency for resource-limited devices such as drones and IoT systems. The research involved data augmentation techniques including crop (0–20% zoom), hue (±20°), saturation (±30%), brightness (±20%), exposure (±15%), and mosaic augmentation. These augmented images were used to train four YOLO nano variants (v5n, v8n, v11n, v12n), which were evaluated using standard metrics: Precision, Recall, F1-Score, and mean Average Precision (mAP). Among the models tested, YOLOv11n with Custom Optuna configuration delivered the highest performance, achieving a 94.6% F1-score and 97.8% mAP@0.5. These results demonstrate that the optimized YOLOv11n model can support accurate and efficient real-time weed detection in household environments, particularly on edge devices with limited hardware capabilities. This makes it a viable solution for practical implementation in precision agriculture and smart gardening.


  • Open Access

    Enhancing Vision Transformer Performance with Rotation Based Augmentation for Classifying Images of Colon Cancer Pathology

    by Rudy Eko Prasetya, M. Arief Soeleman, Farrikh Al Zami, Affandy Affandy, Aris Marjuni, Mohammad Iqbal Saryuddin Assaqty

    Background: In medical imaging, classifying images of colon cancer pathology is still an essential challenge, especially for facilitating early diagnosis and successful intervention. Recently, Vision Transformer (ViT) models have demonstrated great promise for a variety of computer vision tasks, including the classification of medical images. However, the lack of annotated medical datasets and the intrinsic unpredictability of histopathology pictures sometimes restrict their performance. Objective: This study aims to enhance the performance of ViT models in colon cancer pathology classification by introducing a targeted data augmentation strategy, with a particular focus on rotation-based augmentation. Methods: We proposed a data augmentation pipeline that uses controlled changes to improve the number and diversity of training data. Like Rotation, Flip and Geometry are emphasized to replicate the real-world tissue orientation variations that are frequently seen in colon pathology slides. 10,000 JPEG pictures of colon cancer pathology, each with a resolution of 768 x 768 pixels, are used to train the models. We use models trained with and without the suggested augmentation pipeline to compare ViT performance across accuracy, sensitivity, and specificity in order to assess the impact of augmentation. Results: According to study results, rotation-based augmentation enhances ViT performance, achieving up to 99.30% accuracy and 99.50% sensitivity while preserving training times. In real-world pathology settings, where slide orientation varies greatly and can affect categorization consistency, these enhancements are especially pertinent. Conclusion: The proposed rotation-centric data augmentation technique enhances the performance of the ViT model in the classification of images showing colon cancer pathology.