Improving Breast Cancer Screening with Technology

Breast cancer is a big concern, especially as women get older. Health experts suggest that women over 40 or 50 should get an X-ray mammogram every couple of years. In Canada and the USA alone, about 40 million mammograms are done each year. Out of these, around two million are flagged by doctors as suspicious. However, only about 270,000 of those cases turn out to be actual breast cancer. This means that less than 1% of the exams show real cancer. Funny enough, even with all that checking, about 20% of breast cancers slip through the cracks.

To help with getting it right, many facilities are trying double reading (where two doctors look at the same scan) and using artificial intelligence. These efforts seem to help accuracy a bit, but they also make things more expensive and don’t completely fix the problem of false alarms.

Since the 1980s, when more regular screenings started, technology has come a long way. Initially, mammograms were optimized using certain materials for X-ray beams, and special film made the images clearer. Over time, this film technology was replaced by newer digital tech, which allowed better data storage and helped to streamline the whole process. Nowadays, mammograms are usually stored in a format called DICOM, which contains lots of important details about how the images were made.

Innovations in detectors mean that we have cameras that can see better and pick up even the smallest signs of issues. There are also new ways to take pictures of the breast that help doctors see better through overlapping tissue. With thousands of terabytes of images available, computer programs have been created to help spot warning signs.

In the early days, computer programs looked for bright spots in mammograms, which typically meant areas with calcifications that might hint at cancer. But not all cancers show these signs, so researchers started focusing on finding masses, which are different from calcifications. These masses can be trickier to catch because they often don’t look like clear blobs. Instead, they might have fuzzy edges that blend into normal tissue, making them hard to spot, especially in younger women whose breast tissue is often denser and more fibrous.

As computer algorithms get more complex, it becomes trickier to improve how well they work in breast cancer screening. Plus, nobody wants these programs to make things more expensive than they already are. Early studies found that these programs could sometimes make things less efficient because of all the false alarms. While experienced radiologists didn’t see much benefit from using computer help, less experienced ones did. It’s a mixed bag, but many developers keep at it, trying to refine these programs.

However, progress has slowed a bit. Early software was often easier to create because it had a narrower focus. As developers try to tackle tougher problems, there are more chances for things to go wrong. Also, any software used for diagnosis must be approved by medical authorities, which takes time and can uncover a lack of good testing data.

#Existing Datasets: A Review

Many attempts to improve these programs have used some older datasets, like the MIAS and DDSM. There are other datasets out there, but most of them are not easy to access. The popular DDSM dataset, for instance, has way more Abnormal cases than normal ones. This skews the data and can lead to a computer program that doesn’t perform well with real patients.

Also, there are various types of images in these datasets. Some datasets use digitized film images, while others mix different types of digital images. The file formats are all over the place, and many of these datasets don’t even use the standard DICOM format, which is frustrating for anyone trying to license software for medical use.

When it comes to resolution, a program designed to find tiny calcifications may miss them if the images aren’t clear enough. If the resolution is too high, it could get bogged down because of giant files. Lowering quality can help speed things up, but it might lose some important details.

The types of images and the lack of important information make it hard for programmers to come up with reliable software. Many datasets also lack specific details about the equipment used for the mammograms, which can help standardize how the images are processed. The DICOM format is useful because it holds this information in each file’s header, making everything easier.

One big hurdle with computer programs that use deep learning is needing a vast amount of image data for training. The datasets for breast cancer screening often don’t have enough normal and suspicious cases. Some commonly used datasets include:

• DDSM: Contains over 10,000 images, but the number of abnormal cases far exceeds what you would see in a real screening program.

• CBIS-DDSM: A more detailed version of DDSM focused on cancerous images.

• MIAS: A classic dataset with digitized images that have been analyzed for anomalies.

• InBreast: A recent dataset with images from patients, each carefully noted by specialists.

• VinDr: A dataset from Vietnam with thousands of images that also includes scores for breast density and more.

• CMMD: A Chinese dataset that contains a mix of benign and malignant cases.

• RSNA: A large dataset with many images but fewer cancer cases than the others.

• OPTIMAM: A significant UK dataset that tracks interval cancers and biopsy-verified cases.

One interesting dataset is the NL-Breast-Screening (NLBS) dataset, which contains a more realistic mix of cases. The goal was to collect images from a screening program in Newfoundland with proper consent.

In this dataset, all patients diagnosed were confirmed by more tests. Normal cases were verified as free of cancer for at least two years. They collected nearly 27,000 images representing around 6,000 cases. Their dataset is a gem in that it mirrors real-world numbers better than many other datasets.

#The NLBS Dataset and Its Findings

The NLBS dataset includes a mix of normal cases, False Positives, and positive cancer cases. The average age of patients in the positive cancer group was slightly older than in the false positive group, revealing that older women tend to be diagnosed more often. The dataset also includes a variety of images for both sides of the breast and from different views, which is necessary to have a well-rounded analysis.

Even though they have a sizable collection, there’s a worry that there aren’t enough confirmed cases to capture all types of cancer. They plan to keep collecting images to solve this issue. Meanwhile, they can use images from other sources to fill in the gaps, while also keeping in mind that the populations in those datasets may differ from the Canadian population.

The NLBS dataset excludes information about the density of breast tissue but is verified well enough that all normal cases were confirmed free from cancer. The images are in DICOM format, making it easier for researchers to use them.

#Moving Forward: Suggestions and Improvements

In looking at the future of breast cancer screenings and technology, there are a few things to keep in mind:

1. Sensitivity Goals: If we want to find every single positive case, we'll need to analyze normal images more closely to spot any characteristics that might hint at cancer.

2. Use of Advanced Technology: Multiple false positives could mean a shift in how we train our algorithms to make predictions more balanced.

3. Comparing Detection Methods: It’s important to understand if false positives from AI systems are the same as those from radiologists, so we can improve accordingly.

4. Performance Metrics: Reporting how well systems perform through various measures like the AUC (Area Under Curve) and confusion matrices may provide clearer insights.

5. Pre-Processing Techniques: We should look into ways of filtering out irrelevant data to focus on what really matters.

6. Testing Various Features: Trying different features or methods could help identify what works best in detecting breast cancer.

7. Radiation Considerations: It is critical to keep an eye on how much radiation is being used in these procedures.

8. Addressing Fat Content: Predicting the fat content in breast tissues could also help reduce error rates in diagnoses.

9. Incorporating Radiologist Insights: Feeding radiologist summaries into the training data may lead to more accurate models.

10. Follow-Up Procedures: Tracking cancers that show up later and rechecking negative cases could help improve outcomes.

It’s essential to create a solid plan for testing that specifically addresses the problem of data leaks that can skew results.

#Conclusion

Navigating through the landscape of breast cancer screening and artificial intelligence is no small task. While advancements have been made in technology and data collection, several challenges still lie ahead. The ongoing efforts to gather quality datasets and refine diagnostic methods are crucial for improving breast cancer detection and treatment. By keeping an eye on the future and continually adapting to new findings, the ultimate goal remains: to find breast cancer as early as possible and save more lives. Remember, laughter may not cure cancer, but it sure helps lighten the load while we work to beat it!