Early Detection of Pancreatic Cancer through Competitive Lateral-flow Assays and Deep Learning.


With just 5% survival rates, Pancreatic Cancer continues to be one of the most lethal cancers in the world. Recent studies have found out that an increase of polyamine concentration in biofluids correlates with increased cell proliferation and tumorigenesis. In Carcinomas like that of the pancreas, significant up-regulation of the polyamine, Spermine concentration in human saliva can be attested as a marker for Pancreatic Cancer.

The aim of the study was to design a novel strip based assay to scrutinize spermine as a biomarker for the early detection of Pancreatic cancer through saliva. Because of the presence of 40nm carboxyl gold, it shows a color change at the control line, when target spermine present in the saliva binds with anti-spermine gold-conjugated antibody & is detected by the secondary detector probe.  After the test have been conducted, the intensity of the color change would be measured through a Convolution Neural Network based mobile application which would effectively read out the test results & if positive then try to classify in which stage the tumor most likely to be.

The main advantage of this new type of detection scheme is that it is a non-invasive and an affordable test which could easily be conducted by people at home regularly so that the early onset of this cancerous tumor could be detected before it turns to a devastation. Not only this but this novel breakthrough could be used to diagnose other types of diseases & can also revolutionize the field of modern diagnostics.

Question / Proposal

Pancreatic Cancer is one of the most life-threatening cancers in the world. It takes away lives of around 300,000 people worldwide. Currently, there’s an urgent need for the early detection of this disease to increase the survival rates of this deadly cancer.   

Generally, Pancreatic Cancer is diagnosed using:

  • Scans like PET, CT or MRI.  
  • Biopsy or,
  • CA19-9 Blood Test

But according to American College of Radiology, 'no imaging exams should be done unless there is a clear medical benefit', also these scans are too expensive, which is not viable for a normal person to regularly take these scans. Biopsies on the other hand, often lead to confirmatory diagnosis, but a doctor usually recommends a biopsy after the patient has presented with almost all signs for a tumor. As it is an invasive procedure, a patient can't take biopsy tests for general screening. Coming to the CA19-9 blood test, it has a poor specificity & sensitivity of just 65-70%, hence cannot be used as a screening tool for the general population.

Hence all the above diagnosis procedure presents with its own drawbacks which demands a better diagnosis procedure which for being used for regular screening should have the following benifits:

  • Should be Low cost & Efficient.
  • Shouldn't have any potential risks involved.
  • Should be minimal or non-invasive. 

Hence, we decided to create a non-invasive & a highly efficient procedure for detecting Pancreatic Cancer through quantifying the presence of a potential biomarker found in human Saliva using a paper based assay.


We started our research by first studying the different breakthroughs in this field. The first and the most eminent paper we came across many papers on the early detection of Pancreatic Cancer was 'A paper biosensor for early detection of Pancreatic Cancer through carbon Nanotubes & Mesothelin' which detects Mesothelin in serum by inducing a change in electrical properties of the biosensor. After studying the research more thoroughly, we found 3 particular drawbacks in the paper:

  • This paper-based biosensor had an absence of a quantifiable reader which could interpret the differences in the electrical properties as stated and classify the test results.

  • Secondly, Mesothelin, being found in blood, requires a minimally invasive procedure which is generally not agreeable by everyone and is advisable to be conducted in hospitals.

  • Thirdly, Mesothelin was also found in various other cancers such as Lung, Ovarian, Endometrial, Biliary & Gastric Cancers etc. which makes it a low specific biomarker.

So, we started our research, keeping in mind to inexpensively detect a potential biomarker for the purpose. We initially comprehended various biomarkers and chose a few to work on, from the paper: Non-invasive biomarkers for the early detection of Pancreatic Cancer. We surveyed Dr. Devendra Naik, Gastro-surgeon in Balaji Hospital Raipur to know the local statistics. He stated that only CA19-9 (a very insensitive & in specific for Pancreatic Cancer) is used for the purpose other than imaging tests or Biopsies.

We then came across a recent study paper on Elevated polyamines in the saliva of Pancreatic Cancer, which explained about different Polyamines being potential biomarkers for PC & which could be detected in saliva. We finalized our biomarker to be Spermine because of its high sensitivity. This solved the first drawback of making the test totally Non-invasive.

We then decided to design a detection scheme. we first decided to use the most commonly used Immunoassay - ELISA. We fathomed about this assay & learned more from Dr. Ram from Balaji’s Pathology Lab, who demonstrated us about its working. Though ELISA was a significant tool, but was costlier and required various lab equipment, whereas we desired a fast & inexpensive solution.

While studying about other assays, we learned about the working pregnancy test strips, this is where we came across Lateral Flow Assays which are colorimetric and can easily be conducted at home. Combining our biomarker & this test, we prototyped our test. This solved the second drawback of making the test inexpensive and point of care

But then, another challenge approached which was reading the results. Usually done through naked eye which always doesn't lead to precise diagnosis, here we decided to link Convolutional Neural Net. based Mobile App to read out the results. This solved the third drawback of making the test totally quantifiable and highly specific

An intermingling of diverging concepts, ideas & suggestions, we were finally able to design a groundbreaking mechanism which could serve to humanity and help detect this ruinous disease before it destroys its prey.​​​​​​​

Method / Testing and Redesign

Preparation of Strip 

  • The spermine protein was conjugated with the 40nm gold via covalent conjugation. The buffers used were - 5% Sucrose and 10% trehalose.
  • The spermine protein was conjugated with the protein carrier - Bovine Serum Albumin along with the desired buffers.
  • The prepared conjugate was sprayed via an air-jet dispenser onto the conjugate pad keeping the dispense rate 10 μL/cm and OD 10. After spraying, the strip was incubated in an oven for half an hour, followed by overnight drying.
  • The Spermine protein was striped on the test region of the Nitrocellulose Membrane of thickness 200μM ± 5 and flow rate 140 ± 30 seconds via an Iso-flow dispenser with dispense rate 0.5μL/cm and a line width of approximately 1mm. This was also followed by incubation and overnight drying.
  • Lastly, a secondary anti-rabbit polyclonal antibody probe was stripped at the control region with line width and dispense rate the same as above.
  • Finally, after the pad and membrane were ready, they were joined into a backing card via a pressure sensitive adhesive. It was then sealed in a forced-air convention oven for 1 hour in 37-degree centigrade followed by overnight drying at a desiccated environment.

Generation of Images & Training the CNN

  • Six images from different angles were taken on the prepared test strip.
  • Then via Javascript, around 3000 virtual images were generated by randomly plotting lines at the control & the test region as well as randomly adjusting the color and the line width. All the generated images according to their targets were stored under the subfolder named - False / Invalid / True.

This is the image taken after the classifier was retrained

  • Then, a pre-trained Convolutional Neural Network was retrained by setting learning rate 0.03 and training for 1500 epochs by specifying the data directory. The Classifier got trained with 97.4% accuracy and also exported a protobuff file. This was converted to tflite file and added to the mobile application

Saliva Collection Protocol

  • The saliva samples of Pancreatic Cancer were collected from three Local Hospitals - BALCO Cancer Hospital, Sanjeevani Cancer Insitute & Balaji Multispeciality Hospital.
  • The patients were not allowed to eat, drink or consume anything before 1 hour of saliva collection. 
  • The saliva was collected via allowing unstimulated saliva to pass into a sterile vial followed by applying the cover instantly.

Testing the Strip

  • The saliva samples were taken out of the freezer and were centrifuged at 1500g for 15 mins, some saliva samples which appeared viscous were centrifuged at 2500g for another 10 mins.
  • The strip was dipped in the container until the length of the sample pad and kept upright for 3-5 minutes, where the saliva propagated through the nitrocellulose membrane.
  • After a course of 5-8 minutes, the strips displayed the appearance of very diminished lines as the control and the test region. 
  • Then an image was taken an uploaded to the app which classified the results accuately.


Strip Testing 

Test 1 


The first batch of test strips was prepared via the mentioned protocol. These were the following test results:

  • A majority of Saliva took nearly took 7-8 minutes to travel along the nitrocellulose to the wicking pad.

  • Although some saliva samples failed to travel the nitrocellulose membrane.

  • A test line appeared in every test including the positive sample tests which indicated that the concentration of Spermine antibody in the conjugate is higher than the desired threshold.

  • There was an appearance of a very diminished line at the control region when tested for positive samples. This indicated that either the concentration of the detector probe is either too low or the amount of spermine present at the test is very high.


  • Lowering the concentration of Spermine Antibody in the conjugate and Spermine protein in the test region.
  • Using another nitrocellulose membrane with a flow rate of nearly 100 seconds.

Test 2


The second batch of test strips was prepared via the mentioned protocol keeping the above inferences in mind. These were the following test results:

  • There was no significant reduction in the time as the Saliva took nearly took 6-7 minutes which was fair as no samples failed to travel till the end.
  • There were diminished lines at the test region though, at the time with positive samples.


  • The results of this testing were definitely better than the previous testing as it was successful in a majority of the samples. 

  • Some other tweaks have to be made to further increase its sensitivity.

Testing of the Image Classifier

The Image classifier was retrained for 7 times by changing the hyperparameters. The model depicted a minimal 76% accuracy at first but by 7th testing, it was raised up to 97%. 

Test 1


  • The classifier was retrained for 500 epochs and on 1500 images which resulted in an accuracy of 76.8 %.


  • The model's accuracy could be improved by increasing the number of epochs and tuning other hyperparameters.

Test 2


  • The classifier was again retrained for 800 epochs.
  • The number of images was increased to 3705.
  • The default learning rate was changed to 0.002


  • This retrained model had an excellent accuracy and can be used in the app.

The final retrained protobuff file was converted to the tflite format and was imported into the app. When the application was tested, it worked flawlessly.





We have found a novel, non-invasive and a cheap procedure to detect world's one of the most deadliest cancer. This procedure is based on a Lateral Flow Assay system which quantifies the concentration of Spermine in saliva. If the concentration is above the desired threshold than it shows a positive result and if the concentration is below the threshold than shows a negative result.

To test our test strips, we collected saliva samples of freshly diagnosed Pancreatic Cancer (who hadn't been exposed to chemotherapy/radiation or Surgery). The samples were deep freezed for the means of storage. The samples were sent to Medsourse Ozone Biomedicals Delhi, where the tests were conducted. Though the first batch of Strips didn't portrayed excellent results but after variating the concentrations of antibodies and protein in the next batches, the accuracy of the strip to discriminate between Pancreatic Cancer patients vs. Controls. The procedure for improvising the strip is still going on.

On the other hand our Mobile application after being retrained on a dataset of ~ 3700 generated images for 800 epochs depicted an excellent accuracy of ~ 97% in classifying the test results as negatives, positives or invalid. When validated, the classifier accuarately classified all the test images.


  • Only Pancreatic Cancer patients of Stage II to IVB were taken for the test. Other Cancer patients as well as Unsure cases like that of Pancretitis or Diabetic Patients should also be included in the sample size for proper evaluation.
  • The Test reader was trained on a virtual dataset for the virtue of accuracy. To actually make it more accurate, real photos of test strips should be included in the dataset. Also just instead of photos from 6 different angles, variations in the photos must be included to increase the accuracy.

Future Improvements

  • In the future, We are also intending to devise another test strip to detect the presence of microRNA 23a in human saliva, because of its excellent specificity of around 100% in detecting Pancreatic Cancer. It would also evince as a  confirmatory diagnosis for Pancreatic Cancer.

  • We’ll be improving our CNN based application and will be integrating a system of functions to evaluate the growth/dwindle rate of the tumor after the medications through Timely database analysis of the patient through his older test’s results.

  • We’ll also be annexing a class of other diseases which can be detected through the upregulation of the biomarker such as Colorectal Cancer & Colon Cancer. This would help to detect other similar diseases through the same procedure.


This would be a breakthrough in the field of Diagnosis of Cancers as by just a cheap strip based assay, one of the most deadliest cancer of the Pancreas. This could save millions of lives worldwide as well as increase the dismal survival rates of this cancer from just 7%. This innovation will revolutionize the way were think of predicting these deadly mass executioners.

About me

Hello! We are Harsh Agrawal & Anmol Rathi, sophomores at Bhavan's Raipur. We love engrossing in STEM & Research. This endeavor for research arose in us from past two years when two students from our school were qualified to participate in one of the biggest fairs in the world - Intel ISEF. Learning about its passionate summit of some of the youngest researchers committed to solving some of the greatest menaces thrived us to do the same.

Since then, we have dedicated our time to find and solve the greatest threats the world faces today. This is how we came together when we first learned that this disease is a brutal terminator. We committed ourselves to find a way which can prevent people to get under its hold. 

Surveying notable Oncologists & Surgeons or discussing with different professors & mentors, Staying up all night to scrutinize a research paper or Spending the daylight hours trying to tune the Neural Net's hyperparameters to increase its accuracy, nothing has given us more pleasure. 

Our idol is my no one big but Master Shresth Agrawal, the young researcher from our own school, who is also a developer & an alumni of Intel ISEF. He is the one responsible for inspiring us for research pursuits & had helped us in fabricating our ideas from minuscule sparks to massive infernos.

Health & Safety

The test strips were prepared and tested at Medsourse Ozone Biomedicals Delhi under the supervision of Mr. Manan Mittal. Contact Details - manan@ozonebio.com, +91 9810163415. The authentic guidelines were followed throughout the research.

  • Neither the Antibodies, Proteins, Conjugate or any other buffers required for this test had any potential health risks or hazards.
  • Optimal Storage conditions for all the chemicals were followed as per the separate guidelines.
  • For handling lower pH buffers, safety gloves and goggles were worn and proper protocols were followed.

For obtaining patient's saliva sample, a consent was provided to every participant seeking the permission of the patient to be providing the sample on their own will as well as a short information sheet which in brief explains about the study. The safety of study patients was kept in mind. The only discomfort the patients faced was of collecting saliva for 5-10 minutes. 

These samples were approved by the committee of Scientific Research as well as the co-ordinating surgical oncologists of the participating hospitals.

The saliva samples as soon as they were collected were applied a seal and were stored at - 80-degree Celsius for storage at Balaji's Path Labs under the supervision of Dr. Johny David. These samples were shipped to Mr. Manan Mittal at Delhi via an air courier under proper freight conditions.

The testing of Machine Learning Algorithm didn't require for us to follow any special guidelines. It was conducted under the supervision of our project mentor - Mr. Sushil Kumar Pandey, contact: sushilpandey0176@gmail.com, +91 7049636745 and Master Shresth Agrawal, contact: shresthagrawal.31@gmail.com, +91 9425502170.

Bibliography, references, and acknowledgements

Here is the list of websites, journal articles, and websites, we've used in our research:


We take this opportunity to express our profound gratitude and deep regards to Dr. K.D Chawli, Department of Forensic medicine and & Toxicology and Dr. Surya Prakash Dhaneria, Department of Pharmacology from AIIMS Raipur as well as Dr. Devendra Naik, Gastrosurgeon at Balaji Multispeciality Hospital for listening to our project and rendering help wherever required.

We 'd like to thank Dr. Jayesh Sharma, Senior Consultant in the Department of Surgical Oncology, Dr. Shivraj Singh Chauhan, Surgical Specialist of Oncology & Dr. Abdul Danish, Senior Faculty of Emergency at BALCO Cancer Hospital for assisting us in our project as well as approving our study in their hospitals by providing us with PC patients for our research.

We would like to express our regards to Dr. Ankit Thoke, Chemotherapy Specialist & Head of Research & Dr. Yousuf Memon, Director and Surgical Oncologist at Sanjeevani CBCC USA Cancer Hospital Raipur for approving our study and helping by providing Pancreatic Cancer patients as well as assisting us for our project.

We would also like to thank Dr. Ram, Senior Assistant at Balaji Blood Bank for explaining & demonstrating us about ELISA and other Immunoassays & Dr. Jonny David, Senior Assistant in Balaji'sPathology Lab at Balaji Multispeciality Hospital for assisting us in the storage and packaging of saliva samples.

We feel obliged in taking the opportunity to sincerely thank our Project Guide - Mr. Sushil Kumar Pandey who took a keen interest in our project work and guided us all along in the path. A special thanks to Mr. Amitava Ghosh, the Principal of Bharatiya Vidya Bhavan’s R. K. Sarda Vidya Mandir, who made us available all the resources in the school.

We thank Mr. Manan Mittal from Medsource Biomedicals for helping us to redesign and manufacture our test strips, Axiva Scheimbio for providing us the required pads and membranes.

We would also like to thank Rajiv Gandhi Cancer & Advance Research Institute, one of the finest and biggest Cancer Institute in all of Asia for providing the opportunity to present our research at Scientific Committee Meeting at Delhi on 29th January 2019.

We have also been referred to Dr.  Tata Memorial Cancer Institute, another one of the biggest and finest Cancer Institute in Asia by Dr. Aniket Thoke from Sanjeevani Cancer Hospital to present our research in their Scientific Committee meeting. 

We pay our sincere gratitude to our Parents, whom we are greatly indebted for bringing us up with love and encouragement to this stage and have given us different ideas in making this project unique.

Furthermore, we are highly obliged to taking the opportunity to sincerely thank all the Teachers of our school for their generous attitude and friendly behavior. Last but not the least we are thankful to all those who have been always helping and encouraging us throughout our work.