Harnessing Unstructured Data in Radiology: NLP, RadLex & AIM

Introduction

Radiology is the use of imaging to look into the human body and observe disease processes (Chapman et al., 2011). Both diagnosis and treatment can be improved through radiology. Radiologists employ a number of techniques, including CT scans, X-rays, ultrasounds, MRIs, and PET scans, among others (Hong et al., 2013). There are also interventional radiology techniques that are generally minimally invasive but that work well in diagnosing and treating specific ailments (Chapman et al., 2011). However, radiology is severely lacking in one critical area: the mining of unstructured data in order to present a clearer picture of patients' issues and provide more information about what those patients may be facing. A great deal of data is provided within radiology reports, but without collecting and processing that data it remains of no real use to patients or physicians.

The collection and processing of unstructured data found in radiology reports can be difficult and is not without its own pitfalls (Chapman et al., 2011). Several different types of programs can be used to perform this mining successfully. Natural Language Processing (NLP) is one of the most commonly used options for collecting data, but it does not always work well with unstructured data. Errors are frequent when it is applied this way, so it has not been found to be completely reliable. With that in mind, this paper explores NLP, RadLex, and Annotation and Image Markup (AIM) as tools for unstructured data mining in radiology reports, evaluating which method is most effective and how they might be used in conjunction with one another for greater overall success.

Natural Language Processing (NLP) is used to mine unstructured data (Gerstmair et al., 2012; Hong et al., 2013). This concept is based on human-computer interaction and provides a way for computers to learn natural human language in order to process information provided by humans. The more computers understand about language, the more they can process information without barriers (Johnson et al., 1997). That can be highly beneficial in medicine, because it provides doctors, nurses, radiologists, and other medical professionals with more information than they would previously have been able to collect. However, NLP is not without its downsides, which must also be addressed in order to determine whether NLP should be used in radiology and what adjustments could make it more viable.

Historical and Theoretical Background

Generally, complex sets of hand-written rules were used to allow machines to translate language, but in the 1980s programmers began to write complex algorithms that allowed machines to learn and process language (Demner-Fushman, Chapman, & McDonald, 2009; Torres et al., 2012). This was a major breakthrough, and interest in machine translation was renewed. The original algorithms were relatively primitive and not much better than hand-written rules, but they demonstrated that algorithms were possible and that they could function for translation purposes (Chapman et al., 2011; Weiss & Langlotz, 2008). As computing power grew stronger, greater success was seen with translation and data mining, allowing computers to actually "learn" language in ways that had not been possible previously. Algorithms today can be semi-supervised, meaning they can derive some information from other information with which they are supplied (Chapman et al., 2011).

The theory underlying NLP is that computers can be "taught" to translate language in the same way a person can (Reiner, 2009; Torres et al., 2012). Once machines are able to do this, computers will be able to handle a number of tasks currently reserved for humans. That frees human beings for other tasks and results in much faster translations, given that computers are capable of rapid calculations that far exceed what humans can achieve. However, there are issues with this theory that must be considered. The main concern is that the goals of NLP are not entirely realistic (Gerstmair et al., 2012; Weiss & Langlotz, 2008). Computers are not people, and because they do not "think" in the same way human beings do, they can only follow sets of rules and use those rules to process information (Demner-Fushman, Chapman, & McDonald, 2009).

RadLex is another approach to unstructured data mining. Several different methods are currently in use, and the main problem with them is that they are all different from one another. When healthcare organizations use different methods for extracting, categorizing, and storing data, confusion arises when information must be transferred from one organization to another (Gerstmair et al., 2012). RadLex is designed to resolve this through the creation of a single lexicon usable by all healthcare organizations and agencies (Gerstmair et al., 2012). It contains more than 68,000 terms, so it can be applied across the entire field of radiology. DICOM and SNOMED-CT are two of the current standards and lexicons in use, but RadLex is able to work with both in order to unify the experience (Weiss & Langlotz, 2008).

The idea behind RadLex originated with a group of committees formed to find a better way to mine data from radiology reports. The RSNA formed these committees in 2005; they comprised individuals from more than 30 organizations focused on radiology and standards (Chapman et al., 2011). In 2007, six additional committees were formed to continue RadLex development and to ensure that as many terms as possible were included (Chapman et al., 2011). Without that breadth of coverage, RadLex would not have surpassed the other lexicons it was intended to replace, nor would it have been able to integrate those earlier data mining options into one convenient package usable by radiologists everywhere. The underlying theory was to create a solution that would allow all other programs to be merged, and RadLex appears to be succeeding at that goal.

Annotation and Image Markup (AIM) is another approach to mining unstructured data in radiology reports. As the name suggests, it is focused on the images found in reports. However, there is much more to the issue than simply identifying pictures — captions and tags can be attached to these images, supplying readers of the report with a great deal of data that might otherwise be lost (Chapman et al., 2011). AIM is not a new concept; its planning stages go back a number of years. It is not focused on the same types of unstructured data as systems like NLP, however. AIM's focus remains on the images themselves, because failing to provide usable, translatable data alongside these images can cause them to be overlooked (Chapman et al., 2011).

The theory behind AIM for radiology report mining is that a great deal of information is lost within the pictures and images themselves (Chapman et al., 2011). When a report is "read" by a computer program mining its data, the software processes the language within the report text — it makes sense of words, terms, sentences, and other written information. However, the images themselves also contain information, and radiology is fundamentally image-based. Being able to quickly access and interpret these pictures is very important, because they can provide extra insight into the disease or condition of the patient, which can make a significant difference in the speed and accuracy of diagnosis and treatment (Chapman et al., 2011).

Use and Intended Impact

The use of unstructured data mining is varied, but radiology reports represent one of the most important areas in which it is applied. NLP, RadLex, and AIM each play a role in mining the unstructured data contained in these reports — data that can provide meaningful benefit to the radiologist and to the physicians who review the report in order to make a diagnosis and select the best treatment (Chapman et al., 2011; Johnson et al., 1997). The goal is to use software to mine all of that unstructured data and make it available to anyone reading the radiology report (Chapman et al., 2011). This would allow doctors to access information at a glance that they might not otherwise notice, and would also allow them to include more information in patient records accessible to other physicians (Weiss & Langlotz, 2008). Clinical decision-making requires all available information, and unstructured data mining could supply a higher level of detail, leading to better diagnostic success and a greater chance of selecting the right treatments for every patient who undergoes radiology.

There is a risk with this type of data collection, however, because of uncertainty issues that currently exist in its ability to translate correctly and efficiently at all times (Chapman et al., 2011; Demner-Fushman, Chapman, & McDonald, 2009; Do et al., 2013). For unstructured data mining software to be acceptable in a clinical setting, that issue would have to be fully corrected and thoroughly tested so that patients' lives and well-being are not put at risk by incorrect translation. Physicians must be able to trust what they read on a chart or diagnostic report, whether it is produced by another medical professional or translated by a computer (Chapman et al., 2011). With properly executed unstructured data mining, there is an excellent opportunity to collect more information that can help provide patients with the best care possible (Demner-Fushman, Chapman, & McDonald, 2009). As long as the data is collected and translated properly, significant benefits will follow (Chapman et al., 2011).

Using NLP will have an excellent benefit for radiology, provided the translation of any mined unstructured data is correct (Weiss & Langlotz, 2008). The major impact will be on patients themselves, since they are the ones who will truly benefit when more data about their diagnosis and treatment is delivered to their doctors and other medical providers in a structured form. Unstructured data is disorganized and does not lend itself to helping diagnose or treat a patient, regardless of the illness. The structured data in radiology reports is what matters for clinical use, and if NLP can mine unstructured data and accurately convert it into structured data, there will be a significant impact on the value it provides to both doctors and patients (Chapman et al., 2011; Torres et al., 2012). This impact is very important for the field, since it can save lives while also helping doctors diagnose and treat even mild conditions that are causing difficulty for patients (Chapman et al., 2011; Demner-Fushman, Chapman, & McDonald, 2009). If NLP is not used correctly, however, it could have a very negative impact on radiology and other areas of healthcare because of the inaccurate information it might produce.

All of the best features from existing radiology terminology systems are incorporated into RadLex, and the software also fills critical gaps that were missing in other methods used for unstructured data mining of radiology reports. This is vitally important, as the goal is to reach a type or style of software capable of handling unstructured data, structured data, and images alike. While most data mining options are helpful, none of them fully addresses all of the issues faced by those attempting to extract everything a radiology report has to offer (Chapman et al., 2011). RadLex is not perfect, but because it fills the most critical gaps in data mining capability and because it provides a link between all previous options used for creating and mining radiology reports, it represents an excellent choice for unstructured data mining.

The value of AIM is significant for collecting unstructured data as it relates to the pictures included in a radiology report. It is necessary to understand the value of those pictures and to ensure they are conveying the proper information (Chapman et al., 2011). Without images of the patient to help identify the disease or condition, a radiology report does not provide as much value to the physician. That is where AIM comes in, and where it demonstrates the most importance for the medical field. Being able to collect information from pictures in the report and have them be part of a patient record that can be read at any other medical institution is vital to the quality of care a patient receives (Chapman et al., 2011). None of the methods used to mine unstructured data from radiology reports is perfect, but there are many ways in which various methods can work together to produce a highly successful outcome.

Conclusion

Johnson, D. B., Taira, R. K., Cardenas, A. F., & Aberle, D. R. (1997). Extracting information from free text radiology reports. International Journal of Digital Libraries, 1, 297–308.

Reiner, B. (2009). The challenges, opportunities, and imperative of structured reporting in medical imaging. Journal of Digital Imaging, 22(6), 562–568.

Torres, J. S., Quilis, J. D. S., Espert, I. B., & Garcia, V. H. (2012). Improving knowledge management through the support of image examination and data annotation using DICOM structured reporting. Journal of Biomedical Informatics, 45, 1066–1074.

Weiss, D. L., & Langlotz, C. P. (2008). Structured reporting: Patient care enhancement or productivity nightmare? Radiology, 249, 739–747.

Huang, Y., & Lowe, H. J. (2007). A novel hybrid approach to automated negation detection in clinical radiology reports. Journal of the American Medical Information Association, 14, 304–311.

McLoughlin, R. F., So, C. B., & Gray, R. R., et al. (1995). Radiology reports: How much descriptive detail is enough? AJR: American Journal of Roentgenology, 165, 803–806.

Mendonca, E. A., Haas, J., Shagina, L., Larson, E., & Friedman, C. (2005). Extracting information on pneumonia in infants using natural language processing of radiology reports. Journal of Biomedical Informatics, 38, 314–321.

Reiner, B. I., Siegel, E. L., & Knight, N. (2007). Radiology reporting: Past, present, and future: The radiologist perspective. Journal of the American College of Radiology, 5, 313–319.

Rubin, D. L., & Desser, T. S. (2008). A data warehouse for integrating radiologic and pathologic data. Journal of the American College of Radiology, 5, 210–217.