Information extraction from weakly structured radiological reports with natural language queries

OBJECTIVES

Provide physicians and researchers an efficient way to extract information from weakly structured radiology reports with natural language processing (NLP) machine learning models.

METHODS

We evaluate seven different German bidirectional encoder representations from transformers (BERT) models on a dataset of 857,783 unlabeled radiology reports and an annotated reading comprehension dataset in the format of SQuAD 2.0 based on 1223 additional reports.

RESULTS

Continued pre-training of a BERT model on the radiology dataset and a medical online encyclopedia resulted in the most accurate model with an F1-score of 83.97% and an exact match score of 71.63% for answerable questions and 96.01% accuracy in detecting unanswerable questions. Fine-tuning a non-medical model without further pre-training led to the lowest-performing model. The final model proved stable against variation in the formulations of questions and in dealing with questions on topics excluded from the training set.

CONCLUSIONS

General domain BERT models further pre-trained on radiological data achieve high accuracy in answering questions on radiology reports. We propose to integrate our approach into the workflow of medical practitioners and researchers to extract information from radiology reports.

CLINICAL RELEVANCE STATEMENT

By reducing the need for manual searches of radiology reports, radiologists' resources are freed up, which indirectly benefits patients.

KEY POINTS

• BERT models pre-trained on general domain datasets and radiology reports achieve high accuracy (83.97% F1-score) on question-answering for radiology reports. • The best performing model achieves an F1-score of 83.97% for answerable questions and 96.01% accuracy for questions without an answer. • Additional radiology-specific pretraining of all investigated BERT models improves their performance.

Is molecular biology the best alternative for diagnosis of malaria to microscopy? A comparison between microscopy, antigen detection and molecular tests in rural Kenya and urban Tanzania

OBJECTIVE

To assess the agreement of different diagnostic methods for the diagnosis and confirmation of the clinical suspicion of Plasmodium infection in children in Tanzania and Kenya.

METHOD

Blood samples were collected by the finger prick method from 338 children. Blood samples were collected from 338 children with the clinical suspicion of uncomplicated malaria in health clinics in Tanzania and Kenya. The presence of Plasmodium parasites was assessed with microscopy, rapid diagnostic tests (RDTs) and the molecular assays, quantitative nucleic acid sequence based amplification (QT-NASBA) and polymerase chain reaction (PCR). The results were compared and analysed for agreement.

RESULTS

There was a high degree of agreement (88.6-100%) between RDTs or molecular tests and microscopy. In rural Kenya, with a high incidence of malaria cases, the correlation coefficient ranged from 0.94 for RDTs to 0.76 for PCR. In urban Tanzania, where there was a low incidence of cases, R for RDTs was 1.0 but only 0.25 for PCR and 0.33 for NASBA.

CONCLUSION

Malaria is overestimated if the diagnosis is based solely on clinical signs. Therefore, laboratory confirmation is essential. Microscopy is a reliable method in rural areas where malaria is prevalent, but RDTs offer a good alternative with the advantage that it is an easy and rapid method. Molecular tests are more sensitive but difficult to implement in rural areas. In areas with lower incidence, molecular tests detect a significantly higher number of Plasmodium infections than RDTs or microscopy. Although implementation of molecular tools can be difficult, the prospect of an easy and cheap detection system makes them promising tools for the near future.

Three chromosomal rearrangements in neuroblastoma cluster within a 300-kb region on 1p36.1

Deletions in the short arm of chromosome 1 (1p36) and MYCN amplification are common in neuroblastoma. Previously we showed evidence of at least two different neuroblastoma tumor-suppressor loci on 1p. One is associated with MYCN single-copy tumors and maps distal on 1p36.3. A second, more proximal locus maps to 1p36.1 and is deleted in about 90% of neuroblastomas with MYCN amplification. The cell line UHG-NP has the smallest 1p36 deletion of all neuroblastoma cell lines with MYCN amplifications. We assume that the more proximal locus maps within this deletion, close to its proximal border. Here we present the exact localization of the 1p deletion breakpoint of UHG-NP. A 600-kb PAC contig spanning the breakpoint was analyzed for genes and aberrations. Two more neuroblastoma-associated aberrations were mapped within 150 kb of the UHG-NP breakpoint. Within the contig, we identified nine genes expressed in neuroblastoma cells. One of these genes, AML2, maps 200 kb distal to the UHG-NP breakpoint but is expressed only rarely in neuroblastoma and showed no mutations.

Chromosome bands 1p35-36 contain two distinct neuroblastoma tumor suppressor loci, one of which is imprinted

A previous loss of heterozygosity (LOH) study of a series of 91 neuroblastomas suggested that the 1p35-36 region encodes at least two tumor suppressor genes (TSGs) of importance in neuroblastoma development. Here we present the results of a study including 205 neuroblastomas that were analyzed for LOH at chromosome 1 and MYCN amplification. The results corroborate the existence of two TSGs on 1p. Distinct 1p loci seem to be involved in MYCN single copy vs. MYCN amplified neuroblastoma, as these tumors display a different type of shortest region of overlap (SRO). About 15% of MYCN single copy neuroblastomas show 1p deletions of variable length with an SRO of 47 cR at 1p36.3. The lost alleles are preferentially of maternal origin (P = 0.0002), suggesting parental imprinting of the locus. MYCN amplified neuroblastomas have a contrasting pattern of 1p loss. These tumors display much larger deletions of at least 89 cR comprising the region from 1p36.1 to the telomere. LOH of 1p is detected in 86% of the cases. The lost alleles are of random parental origin, suggesting inactivation of a non-imprinted TSG.

The MEIS1 oncogene is highly expressed in neuroblastoma and amplified in cell line IMR32

Neuroblastoma is an embryonal tumor originating from neural crest-derived cells. Here we present the serendipitous cloning of amplified sequences of chromosome 2p15 in neuroblastoma cell line IMR32. The amplified region was analyzed for oncogene activation using a SAGE (serial analysis of gene expression) library of IMR32. SAGE permits a quantitative analysis of all transcripts of a tissue or cell line. The expression of genes and ESTs mapping within a 30-cR region covering the amplicon was compared to 4 additional SAGE libraries of neuroblastomas and 12 SAGE libraries of other tissues in the CGAP databases. The IMR32 SAGE database revealed increased expression of the MEIS1 oncogene, whereas other SAGE libraries showed little or no MEIS1 expression. MEIS1 turned out to be highly amplified and overexpressed in IMR32. Analysis of 24 neuroblastoma cell lines and 22 tumors showed high-level expression in about 25% of the cases. The MEIS1 homeobox protein forms a complex with the HOXA9 and PBX proteins that are implicated in human leukemia. MEIS1 is a target of retroviral insertion in murine leukemia. This is the first report of a MEIS1 amplification and high expression levels in human cancer and the first time that identification of a candidate target of amplification is facilitated by high-throughput mRNA expression profiling.

An integrated 5-Mb physical, genetic, and radiation hybrid map of a 1p36.1 region implicated in neuroblastoma pathogenesis

Common genetic aberrations of neuroblastoma are deletions of the short arm of chromosome 1 (1p36) and MYCN amplification. Our deletion analysis of 25 tumor cell lines and 171 tumors strongly suggests that 1p harbors several tumor suppressor loci. Distinct loci are involved in MYCN single-copy versus MYCN-amplified neuroblastoma. Deletions in MYCN single-copy tumors have a shortest region of overlap (SRO) of 20 cM at 1p36.3. MYCN-amplified tumors have large deletions with an SRO of about 60 cM, from 1p36.1 to the telomere. This SRO is defined by D1S7 (1p36.1), which was the most distal locus retained. Therefore, a suppressor gene associated with MYCN-amplified tumors probably maps within a few megabases distal of D1S7. In order to map this locus, we further refined this SRO. We mapped the breakpoint of the MYCN-amplified neuroblastoma with the smallest 1p deletion between 56.6 and 57.2 cM from 1pter. Pulsed-field gel electrophoresis and radiation hybrid mapping were used to construct a 5-Mb physical map of this region. The map includes the region from 82.73 till 92.89 cR from 1pter. About half of it was isolated in P1 and PAC clones. The region harbors the genes FGR, SLC9A1, HMG17, EXTL1, AML2, RH, OP18, four ESTs, and a newly identified gene with a transcript size of approximately 7 Kb. Several of the mapped genes have a putative role in cell growth, differentiation, and morphogenesis. Genes Chromosomes Cancer 27:143-152, 2000.

Refined physical mapping and genomic structure of the EXTL1 gene

Recently, the EXTL1 gene, a member of the EXT tumor suppressor gene family, has been mapped to 1p36, a chromosome region which is frequently implicated in a wide variety of malignancies, including breast carcinoma, colorectal cancer and neuroblastoma. In this study, we show that the EXTL1 gene is located between the genetic markers D1S511 and D1S234 within 200 kb of the LAP18 gene on chromosome 1p36. 1, a region which has been proposed to harbor a tumor suppressor gene implicated in MYCN-amplified neuroblastomas. In addition, we determined the genomic structure of the EXTL1 gene, revealing that the EXTL1 coding sequence spans 11 exons within a 50-kb region.

Stressor-specific enhancement of hsp induction by low doses of stressors in conditions of self- and cross-sensitization

In this paper, the pattern of induction of heat shock proteins (hsps) was studied in cultured Reuber H35 rat hepatoma cells by sequential application of different stressors. We analyzed whether a specific stress condition is able to induce an enhanced sensitivity to a subsequent application of a low dose of either the same or another stressor (self-sensitization and cross-sensitization, respectively). As a measure of sensitization, the stimulation of hsp induction was employed. Three different stressor conditions (heat shock, sodium arsenite and cadmium chloride) were used in doses which exerted a similar impact on overall protein synthesis. A synergistic effect in induction of the synthesis of various hsps was observed when a high stressor dose was followed by an 8-h incubation in a lower stressor dose in both self- and cross-sensitization experiments. The low-dose conditions used as second treatments did not induce any responses in non-pretreated cells. Studies in cultured cells have demonstrated stressor-specific hsp induction patterns. In this study we analyzed whether the pattern of hsps induced by the low-dose condition is characteristic for the first sensitizing stressor or for the secondary stressor applied in a low dose. The pattern of hsps which was induced above the level of the high-dose effect, due to the incubation with the secondary applied low-dose condition, was found to be characteristic for the secondary stressor and not for the sensitizing primary treatment. These results are of importance for an improved understanding of the regulation of heat shock protein synthesis in conditions of self- and cross-sensitization, as well as for a proper use of hsps as biomarkers of exposure to environmental stress.

Is heat shock protein re-induction during tolerance related to the stressor-specific induction of heat shock proteins?

The existence of stressor-specific induction programs of heat shock proteins (hsps) leads us to analyze the possible occurrence of a stressor-specific tolerance induced by either heat shock, arsenite, or cadmium. As a measure of this tolerance re-induction of hsps was studied. In this paper, we tested whether the refractory state is either valid for each specific hsp (implying independent regulation of every member of the heat shock protein family) or extends from small subsets of the hsp-family to even larger groups of proteins (indicating a more common denominator in their regulation). (re-)induction of hsps does not seem to be regulated at the level of each individual hsp since differences in induced synthesis of hsps between two stressor conditions are not supplemented systematically upon the sequential application of the two stressors. The most notable example in this respect is hsp60. A pretreatment with cadmium, which hardly induces synthesis of this hsp, does induce a tolerance to (re)-induction by heat shock, which normally induces hsp60. This suggests the existence of a more common denominator regulating the coordinate expression of at least some hsps. From our data we conclude that the degree, but not the pattern, of hsp re-induction is influenced by the type of stressor used in the pretreatment. The pattern of hsps induced by a secondary applied stressor still shows most of its stressor-specificity and seems to be independent of any pretreatment. The possible implications of stressor-specificity are discussed.

E-cadherin expression and in vitro invasion of canine mammary tumor cells

E-cadherin is considered to be an invasion suppressor molecule. We have studied the expression and function of E-cadherin in three cell lines derived from a dog mammary tumor, namely SH15, SH24, and SH27. In monolayer culture the cell lines can be distinguished by their morphotype: epithelioid (SH15), fibroblast-like (SH24) and intermediate type (SH27). SH27 was unable to form colonies in collagen gel in contrast to SH15 and SH24. All three cell lines expressed the E-cadherin antigen, as evident from immunocytochemistry, and alpha-, beta-, and gamma-catenins as evident from immunoprecipitation with E-cadherin antibody. Only SH27 showed E-cadherin-dependent aggregation, and little invasion into collagen type 1 gels, in contrast to SH15 and SH24 cells. However, in the precultured embryonic chick heart assay all three cell lines were invasive, demonstrating that invasion depends upon the microenvironment. We assume that in the embryonic chick heart, factors were present or were induced by the SH27 cells, interfering with the function of E-cadherin.