Hello-

 1.) It might help to know the dcm2nii version and operating system you are using. This is listed when you run the program, for example on my OS X computer I am using "dcm2nii :: 4AUGUST2014 32bit BSD License"

2.) I am not sure if you are using dcm2nii (which comes with MRIcron) or the newer dcm2niix (which comes with MRIcroGL). I would recommend you try out dcm2niix - it was designed from the ground up for large datasets.

3.) The standard NIfTI 16-bit data type is 16-bit SIGNED integers - this can losslessly encode values from -32768..+32767. This is typically fine, as most classic MRI systems use 12-bit ADC (range 0..4096). However, it is clear that your software is set up to convert to 16-bit UNSIGNED (0..65535). This is really way beyond the SNR expected for EPI data, but it poses a problem for true lossless conversion. In this situation my software provides you with two options:
 a.) save as unsigned 16-bit integers (UINT16): this keeps the raw data from your scanner unchanged. However, this is a legal but atypical NIfTI data format so subsequent tools may fail.
 b.) save as 32-bit floating point (FLOAT32): this is a standard format that has ample precision. However, the files will require twice as much disk space and will probably require more time to process.
 So I think the choice depends on what tools you will use for subsequent analysis. 

 4.) In general, if data in different folders is from different image series (e.g. either a different individual or a different sequence) you could process on folder at a time. This will dramatically reduce the memory usage on your machine. In addition, I would suggest you get a 64-bit version of dcm2niix newer than June 12, 2015: this will use dramatically less memory than other tools for conversion. You can get the source code and some recent compilations from
   https://github.com/neurolabusc/dcm2niix 

5.) If all else fails you can provide me with a sample dataset to evaluate.