A5084801169- AI in cancer detection
- Radiomics and Machine Learning in Medical Imaging
- Global Cancer Incidence and Screening
- Topic Modeling
- Biomedical Text Mining and Ontologies
- Adversarial Robustness in Machine Learning
- Privacy-Preserving Technologies in Data
- Lung Cancer Diagnosis and Treatment
- Lung Cancer Treatments and Mutations
- Artificial Intelligence in Healthcare and Education
- Breast Lesions and Carcinomas
- Breast Cancer Treatment Studies
- Natural Language Processing Techniques
- Generative Adversarial Networks and Image Synthesis
- Web Data Mining and Analysis
- SARS-CoV-2 and COVID-19 Research
- Colorectal Cancer Screening and Detection
- Dermatological and COVID-19 studies
- Cancer Immunotherapy and Biomarkers
- Machine Learning in Healthcare
- COVID-19 Clinical Research Studies
- Data Quality and Management
- Digital Radiography and Breast Imaging
- Pancreatic and Hepatic Oncology Research
- Breast Implant and Reconstruction
University of California, Berkeley
2023-2025
Maternity and Children's Hospital
2024
Massachusetts Institute of Technology
2017-2023
Berkeley College
2023
The University of Texas MD Anderson Cancer Center
2021
The Gordon Hospital
2021
Tecnológico de Monterrey
2021
Hospital Zambrano Hellion
2021
IIT@MIT
2021
Harvard University
2018-2019
Background Mammographic density improves the accuracy of breast cancer risk models. However, use is limited by subjective assessment, variation across radiologists, and restricted data. A mammography-based deep learning (DL) model may provide more accurate prediction. Purpose To develop a DL that than established clinical Materials Methods This retrospective study included 88 994 consecutive screening mammograms in 39 571 women between January 1, 2009, December 31, 2012. For each patient,...
Purpose To develop a deep learning (DL) algorithm to assess mammographic breast density. Materials and Methods In this retrospective study, convolutional neural network was trained Breast Imaging Reporting Data System (BI-RADS) density based on the original interpretation by an experienced radiologist of 41 479 digital screening mammograms obtained in 27 684 women from January 2009 May 2011. The resulting tested held-out test set 8677 5741 women. addition, five radiologists performed reader...
Background Recent deep learning (DL) approaches have shown promise in improving sensitivity but not addressed limitations radiologist specificity or efficiency. Purpose To develop a DL model to triage portion of mammograms as cancer free, performance and workflow Materials Methods In this retrospective study, 223 109 consecutive screening performed 66 661 women from January 2009 December 2016 were collected with outcomes obtained through linkage regional tumor registry. This cohort was split...
An algorithm to predict breast cancer risk, Mirai, outperforms clinical risk models across test cohorts from the United States, Sweden, and Taiwan.
PURPOSE Low-dose computed tomography (LDCT) for lung cancer screening is effective, although most eligible people are not being screened. Tools that provide personalized future risk assessment could focus approaches toward those likely to benefit. We hypothesized a deep learning model assessing the entire volumetric LDCT data be built predict individual without requiring additional demographic or clinical data. METHODS developed called Sybil using LDCTs from National Lung Screening Trial...
Accurate risk assessment is essential for the success of population screening programs in breast cancer. Models with high sensitivity and specificity would enable to target more elaborate efforts high-risk populations, while minimizing overtreatment rest. Artificial intelligence (AI)-based models have demonstrated a significant advance over used today clinical practice. However, responsible deployment novel AI requires careful validation across diverse populations. To this end, we validate...
Most successful information extraction systems operate with access to a large collection of documents. In this work, we explore the task acquiring and incorporating external evidence improve accuracy in domains where amount training data is scarce. This process entails issuing search queries, from new sources reconciliation extracted values, which are repeated until sufficient collected. We approach problem using reinforcement learning framework our model learns select optimal actions based...
Recent advancements in text-to-image diffusion models have yielded impressive results generating realistic and diverse images. However, these still struggle with complex prompts, such as those that involve numeracy spatial reasoning. This work proposes to enhance prompt understanding capabilities models. Our method leverages a pretrained large language model (LLM) for grounded generation novel two-stage process. In the first stage, LLM generates scene layout comprises captioned bounding...
Abstract High-throughput phenotypic screening has historically relied on manually selected features, limiting our ability to capture complex cellular processes, particularly neuronal activity dynamics. While recent advances in self-supervised learning have revolutionized the study morphology and transcriptomics, dynamic processes remained challenging phenotypically profile. To address this limitation, we developed Plexus, a novel model specifically designed quantify network-level Unlike...
Artificial intelligence (AI) programs in radiology typically provide a numeric score for each case that correlates with the underlying pathology. However, these scores are not readily interpretable by themselves. To address this, we propose improving interpretability providing False Discovery Rate (FDR) and Omission (FOR) corresponding threshold. Using an open-source AI program breast cancer, estimated FDR FOR across range of using data from 130,712 digital screening mammograms, which 907...
Deep Learning Model to Assess Cancer Risk on the Basis of a Breast MR Image AloneTally Portnoi1, Adam Yala1, Tal Schuster1, Regina Barzilay1, Brian Dontchos2,3, Leslie Lamb2,3 and Constance Lehman2,3Audio Available | Share
We propose a novel approach to conformal prediction for generative language models (LMs). Standard produces sets -- in place of single predictions that have rigorous, statistical performance guarantees. LM responses are typically sampled from the model's predicted distribution over large, combinatorial output space natural language. Translating this process prediction, we calibrate stopping rule sampling different outputs get added growing set candidates until confident is sufficient. Since...
In this work, we re-examine inter-patch dependencies in the decoding mechanism of masked autoencoders (MAE). We decompose for patch reconstruction MAE into self-attention and cross-attention. Our investigations suggest that between mask patches is not essential learning good representations. To end, propose a novel pretraining framework: Cross-Attention Masked Autoencoders (CrossMAE). CrossMAE's decoder leverages only cross-attention visible tokens, with no degradation downstream...
1 Abstract Purpose Extracting information from Electronic Medical Record is a time-consuming and expensive process when done manually. Rule-based machine learning techniques are two approaches to solving this problem. In study, we trained model on pathology reports extract pertinent tumor characteristics, which enabled us create large database of attribute searchable reports. This can be used identify cohorts patients with characteristics interest. Methods We collected total 91,505 breast...
Natural language processing (NLP) techniques have been adopted to reduce the curation costs of electronic health records. However, studies questioned whether such can be applied data from previously unseen institutions. We investigated performance a common neural NLP algorithm on both known and heldout (ie, institutions whose were withheld training set only used for testing) hospitals. also explored how diversity in affects system's generalization ability.We collected 24,881 breast pathology...