160 likes | 287 Vues
This project focuses on retrieving current information related to lung abnormalities from free-text radiological reports. It addresses crucial challenges such as locating findings, distinguishing recent findings from past ones, resolving co-references, handling negations, and assessing changes over time. Using advanced techniques like grammar parsing and pattern matching, it aims to extract precise location and size data for various lung lesions. The project demonstrates a current accuracy of 80.82% in classifying findings and outlines future work to enhance data collection and co-reference resolution techniques.
E N D
Dmitriy Zinovev CSC 594 Final Presentation
Project Description • Retrieve current information related to lung abnormalities from the free-text radiological reports. Finding Location Size Within the anterior segment of the right upper lobe, there is a 1.2 x 1.7 cm peripheral lesion which comes close to the pleural surface.
Issues • Locating findings • Distinguishing recent findings from past ones • Resolving co-references • Taking negation into account • Determining whether a change has been noticed since last exam
Locating findings • …decrease in size of a small spiculated mass in the anterior… • …nodular opacity is seen… • …a rounded nodule in the right lower lobe… • …tiny 2 mm left upper lobe micronodule… • …is a bilobed focus, measuring 11 x 9 mm… • Create a dictionary of terms • Search “findings” portion of the report for the matches
Locating findings • Extract location & size information associated with finding. • Search for location terms within the sentence using designated “location” dictionary. • Search for size within the sentence using pattern matching. [0-9]+ ,? [0-9]+ (X? [0-9]+ ,? [0-9]+)? _? (c|m)m • Use grammar parser to determine whether the extracted info refers to the finding
Distinguishing recent findings from past ones On 6/9/05, this right paratrachealmass measured 5.6 x 6.3 cm and currently measures 5.4 x 5.6 cm (series 2, image 18). • Use grammar parses to check whether the verb associated with the subject has past tense
Resolving co-references • In the left upper lobe, a triangular nodule is present on series 5 image 5. It also was present on the study from 10/17/02. It appears slightly more triangular in shape on the current study but has not changed significantly between exams.
Resolving co-references • The majority of the mass is of low attenuation centrally, with a thin rim of tissue present around the periphery. It currently measures 42 x 37 mm, decreased from 49 mm on the previous study.
Resolving co-references • If information of interest has not been found in the same sentence where finding was located, check if the subject of next matches with one of keywords. (This|It)
Taking negation into account • Noadrenal masses are identified. • No new pulmonary parenchymalabnormalities are identified • There are no pulmonary nodules or masses • There are no focal areas of pulmonary consolidation
Determining whether a change has been noticed since last exam • A 4 mm left lower lobe micronodule is again seen, which is unchanged in size and appearance • Both were present on a remote CT dating back to 10/17/02 and are stable.
Accuracy • Percentage of correctly classified findings identified as current: • 80.82%
Future work • Obtain more data • Build more sophisticated co-reference resolution tool • Generalize the rest of techniques